AI in Research & FullMag — Practical Workflows and Micromagnetic Simulation

AI · MODEL NAMESA1b

Sol, Luna, Astra, Terra: a visual introduction

User-supplied illustration associating Sol, Luna, Astra and Terra with celestial imagery

Illustration, not an official capability ranking. Use the next slides for model selection and reasoning settings.

AI · PRACTICAL RESEARCH WORKFLOWA2

Know the model families — then pick the task

A provider, a model, a reasoning setting and a subscription are different choices.

Provider	Efficient / focused	Everyday work	Demanding work
OpenAI	GPT-6 Luna MAX Extraction, repeated edits	GPT-6.1 Terra: visual work / writing GPT-6.1 Sol: analysis / coding	GPT-6 Astra Difficult synthesis and decisions
Anthropic	Haiku 5.5 Fast, bounded tasks	Sonnet 5.5 · general work Opus 5.5 · demanding agent work	Fable 5.1 Frontier coding / agentic work
Google	Gemini 3.5 Flash-Lite · routine intake	Gemini 3.8 Flash · general agent work	Gemini 4 Argon · flagship; 3.1 Pro Preview · preview

Starting recommendations — validate on your own papers, scripts and numerical checks.

AI · PRACTICAL RESEARCH WORKFLOWA2b

Choose the model — then choose how hard it thinks

Reasoning effort trades response time and token use for deeper analysis.

Claude · model → Effort

Choose the model, then adjust Effort.
The screenshot shows High.

ChatGPT / Work · model + reasoning

OpenAI model menu; reasoning setting is not visible in this screenshot

Luna: MAX. Others: start with the default.

Where available: Power → Advanced to set model and reasoning effort.

Move toward Smarter for deeper analysis; Faster for routine work.

This capture shows model selection only. Controls vary by client and plan.

Low / fasterDOI extraction, formatting, small edits

Default / mediumRoutine analysis, scripts, manuscript edits

Higher effortDerivations, difficult bugs, conflicting evidence

For Luna, always choose MAX. For other models, adjust effort to the task and check quality, time and usage.

Model selection and reasoning effort are distinct. These task mappings are practical starting heuristics, not equivalent capability levels across vendors. OpenAI advises starting at the default and increasing effort for complex planning/analysis; desktop/Work can expose Power and Advanced controls. The supplied OpenAI screenshot only demonstrates model choice, not the location or selected effort. Claude exposes Effort in the model menu; higher effort takes longer and consumes more tokens/usage. Thinking visibility and effort are separate. Do not present API effort controls as identical to every app UI. The screenshots show this user’s account, not guaranteed availability for the audience. Ask for concise output separately from reasoning effort. Check sources, units and tests regardless of effort.

Presenter recommendation: always use GPT-6 Luna at MAX reasoning. This is a workflow preference, not a claim that benchmarks used MAX or that all models should use maximum effort.

OPTIONAL · AX3 · SUPPORTING DETAILSkip extras →

Chat controls and Work model names are not identical

Select the actual model and effort; a prompt saying “think harder” is not a settings change.

Surface	What to choose	Practical use
ChatGPT Chat	GPT-6 thinking slider; eligible Pro uses Astra	Bounded analysis, explanation and research
ChatGPT Work / Codex	Astra / Sol / Terra: choose effort; Luna: always MAX	Files, applications and project execution
Claude / Claude Code	Available family and effort in the account selector	Same task contract; different tool access
Gemini app / API	App mode or exact API model ID	Do not equate a subscription name with a model

Pro subscription ≠ Pro model / effort ≠ Deep Research. Availability varies with plan, client and rollout.

AI · PRACTICAL RESEARCH WORKFLOWA3

Which model should a researcher use?

Route the work by difficulty and cost of error.

Research task	Starting choice	Escalate when…
Extract DOI, units or table fields	Luna MAX / Haiku / Flash-Lite	The source is ambiguous or multimodal
Rewrite a paragraph; inspect figures	Terra / Sonnet / Flash	Meaning, evidence or notation conflicts
Debug analysis; compare methods	Sol / Opus / Gemini Pro	The cause spans files or assumptions
Derive a model; challenge a research plan	Astra / Fable / Argon; bounded strong reasoning	An independent check still finds contradictions

One workflow can use several models: strong planner → economical executor → focused independent check.

AI · PRACTICAL RESEARCH WORKFLOWA4

Set your research defaults once

Personality controls tone. Instructions define the working rules.

Where to set them

ChatGPT: Settings → Personalization → Base style and tone / custom instructions.

Claude: profile preferences, response styles and project instructions.

Project: keep symbols, methods and constraints with the project.

A useful researcher preset

Use concise scientific English and SI units. Preserve citation keys and uncertainty. Separate evidence from interpretation. Never invent references or measurements. Flag missing sources and check limiting cases.

AI · PRACTICAL RESEARCH WORKFLOWA4b

My usage: what would the API equivalent cost?

A transparent estimate using GPT-6.1 Sol — the model I use most.

268.2 billion tokens

Displayed account total · mostly Sol, according to the author.

What if these were GPT-6.1 Sol API tokens?

Illustrative token mix	Estimated cost
90% input / 10% output · no cache	$751k
90% input / 10% output · 90% input cache hits	$350k
99% input / 1% output · 90% input cache hits	$117k

USD / 1M: input $2 · output $10 · cache read $0.10 · 30-minute cache write $2.50.

Estimates, not a bill. Standard tier; ≤272K input/request. Cached variants count misses as cache writes. Excludes tools and tax.

Official API pricing · checked 28 Oct 2026

The profile displays 268.2 mld (billion) total tokens, 6.9 billion maximum tokens, a 73h39m longest task and an 81-day streak. Only the 268.2 billion total is used. The user says most usage is Sol. These are counterfactual all-GPT-6.1-Sol equivalents at current rates, not historical invoicing or measured savings. The profile does not reveal models, input/output/cache split, actual billing semantics or token dates. Same token count does not mean equal quality on a different model. Formula without cache: T/1e6*(input_share*2+output_share*10). With 90% input cache hits: T/1e6*(input_share*0.1*2.5+input_share*0.9*0.1+output_share*10). All misses in cached scenarios are assumed 30-minute cache writes. Output share must include billed reasoning tokens. Exact scenario totals: 90% input / 10% output · no cache: $750,960.00, 90% input / 10% output · 90% input cache hits: $350,269.20, 99% input / 1% output · 90% input cache hits: $117,096.12. No statement that these scenarios bound the true cost. A 12-hour cache, other models, longer requests, premium tiers, tools or different output shares change the result.

OPTIONAL · AX1 · SUPPORTING DETAILSkip extras →

API prices: OpenAI and Anthropic

USD per 1M tokens · standard direct API · checked 28 October 2026

Model	Input	Output	Cached input
GPT-6 Luna	$0.10	$0.50	$0.01
GPT-6.1 Terra	$1	$5	$0.05
GPT-6.1 Sol	$2	$10	$0.10
GPT-6 Astra	$10	$50	$1
Claude Haiku 5.5	$0.10	$0.50	$0.01
Claude Sonnet 5.5	$2	$10	$0.10
Claude Opus 5.5	$4	$20	$0.20
Claude Fable 5.1	$10	$50	$0.25

OpenAI: ≤272K input tokens. Haiku: ≤100K total prompt; above this, input/output becomes $0.50/$2.50. Tools, cache writes and other tiers are extra.

OPTIONAL · AX2 · SUPPORTING DETAILSkip extras →

API prices: the current Gemini lineup

Current paid Developer API text rates · promotional changes are explicit.

Model	Input / output per 1M	Status / condition
Gemini 4 Argon	$2 / $10	Introductory through 30 Nov 2026; then $4 / $20
Gemini 3.8 Flash	$0.75 / $3.75	Introductory through 31 Dec 2026; then $1.50 / $7.50
Gemini 3.5 Flash-Lite	$0.30 / $2.50	Stable
Gemini 3.1 Pro Preview	$2 / $12	≤200K prompt; above: $4 / $18
Gemini 3.1 Flash-Lite	$0.25 / $1.50	Earlier stable option; text / image / video

Thinking is included in billed output. Caching, storage and grounding have their own accounting. App subscriptions use separate access rules.

AI · PRACTICAL RESEARCH WORKFLOWA5

Use memory for continuity — files for exact facts

Carry preferences forward without turning recollection into evidence.

Worth remembering

Your field, preferred language and notation
How the group reviews a manuscript
Recurring project conventions

Ask: “What do you remember about this project?” Correct stale details.

Keep in versioned files

Material parameters and units
Accepted derivations and source passages
Current decisions and unresolved questions

Use: CONTEXT.md, DECISIONS.md, references.bib.

ChatGPT web memory and local Codex memory are separate stores. Do not assume automatic transfer to VS Code.

AI · PRACTICAL RESEARCH WORKFLOWA6

A skill is a reusable laboratory procedure

Instructions + reference material + optional scripts — loaded when relevant.

Example: manuscript claim audit

Read the manuscript and cited sources.
Extract claims with line or section locations.
Match each claim to a source passage.
Return supported / overstated / unsupported.
Preserve the original; propose a diff.

What goes in the skill

claim-audit/
  SKILL.md
  references/checklist.md
  assets/audit-template.csv
  scripts/check_citations.py

Different roles: skill = procedure; project = context; connector = access; model = reasoning engine.

Test the procedure on one known manuscript before reusing it across the group.

OPTIONAL · AX8 · SUPPORTING DETAILSkip extras →

Where should a reusable instruction live?

Keep style, memory, project rules and skills separate.

Need	Chat / application	Local project
Preferred language and tone	Personal preferences / style	Personal agent instructions
Remember recurring context	Memory, if enabled and in scope	Codex host memory / Claude Code auto memory
Enforce project conventions	Project instructions and knowledge	AGENTS.md / CLAUDE.md; checked-in methods
Repeat a procedure	Supported skill / custom assistant / Gem	SKILL.md with references and scripts

Chat history, local memory and project files are not automatically synchronized across products.

AI · PRACTICAL EXAMPLESA6b

A small skill kit for a research group

Use one skill for a specific stage; these are available examples, not universal built-ins.

Skill	When to invoke it	Useful result
idea-refine · brainstorming ↗	An idea is still vague	Alternative directions + assumptions + smallest test
planning-and-task-breakdown ↗	The direction is chosen	Ordered tasks, dependencies and acceptance checks
debugging-and-error-recovery ↗	A script or numerical result is wrong	Reproduction → cause → targeted fix → verification
code-review-and-quality ↗	Before trusting an analysis change	Prioritized findings with file locations and evidence
source-driven-development ↗	Using an unfamiliar library / API	Implementation grounded in official documentation
claim-audit · included template ↗	Checking a manuscript draft	Claim–source table; unsupported claims flagged

Search skills.sh ↗ · How to find, inspect and install →

OPTIONAL · AX12 · SUPPORTING DETAILSkip extras →

Find a skill by the task you need done

Discover → inspect the source → install selectively → test on a small example.

1 · Search and inspect

Browse skills.sh ↗ or ask an available find-skills skill ↗.

npx skills find brainstorming
npx skills find literature review

Read SKILL.md, scripts, dependencies and license. Check author and maintenance. Popularity alone is not a quality guarantee.

2 · Choose one skill

npx skills add addyosmani/agent-skills --list

npx skills add addyosmani/agent-skills --skill idea-refine

Run in your project; select the intended agent. Then ask it to list available skills and use the exact installed name.

CLI setup and current commands ↗

“Find a skill for checking manuscript claims against source PDFs. Compare three options by source, workflow, dependencies and output. Show links before installing anything.”

AI · PRACTICAL EXAMPLESA6c

Brainstorming: turn an idea into a testable direction

Example: idea-refine ↗ · Alternative: Superpowers brainstorming ↗

Prompt to copy

Use the available idea-refine skill. We want to study spin-wave confinement in a YIG waveguide. We have four weeks, existing simulation tools and no new measurements.

Ask up to three clarifying questions. Propose five distinct directions. Compare scientific value, feasibility and the hardest assumption. After I choose, propose the smallest discriminating test.

What the skill adds

Opens alternatives before committing.
Challenges assumptions and weak ideas.
Turns the selected idea into a short brief.

Expected output: comparison table → chosen hypothesis → test and failure criterion.

Novelty is a hypothesis until checked against literature.

AI · PRACTICAL RESEARCH WORKFLOWA7

Chat → research package → Work / Code

Use deep analysis to produce a handoff that another agent can execute.

1 · Explore in Chat

Use Pro / strong reasoning for a bounded derivation. Use Deep Research for source discovery and synthesis.

→

2 · Export the evidence

brief.md
methods.md
evidence.csv
references.bib
open_questions.md
report.pdf

→

3 · Continue with files

Download a ZIP, inspect it, open the folder in VS Code or attach it to Work.

Implement one specific analysis or revision.

“Read brief.md and methods.md. Implement only Figure 2. Preserve inputs. Run unit and limit checks; return the diff and reproduction command.”

OPTIONAL · AX7 · SUPPORTING DETAILSkip extras →

A portable research handoff

Download the templates; replace the examples with your verified project facts.

Before leaving Chat

brief.md
accepted_methods.md
evidence.csv
references.bib
open_questions.md
README.md

PDF for reading · editable sources for execution · ZIP for transport.

First prompt in VS Code

Read the brief and source manifest. List missing inputs. Implement only the agreed analysis. Preserve raw data. Return a diff, reproduction command and checks actually run.

Download the researcher starter pack

AI · PRACTICAL RESEARCH WORKFLOWA8

VS Code is a research workspace, not just a code editor

Put the manuscript, data, scripts, figures and agent in one reviewable project.

User-provided Fullmag VS Code workspace with Claude Code and Codex

A practical loop

Open the project folder.
Reference only relevant files.
Ask for one bounded change.
Inspect the diff and test output.

Use Codex or Claude Code; the extensions and memory scopes are distinct.

Optional: Claude Code + Overleaf tutorial →

OPTIONAL · AX10 · SUPPORTING DETAILSkip extras →

Prompt: implement one analysis in VS Code

Attach the raw data, units, method and one known reference case.

Read brief.md, dispersion.csv and units.md. Write a script that plots frequency in GHz against wave number in rad/µm. Preserve the raw CSV. Use only the analytical expression supplied in methods.md; flag undefined parameters before implementation.

Return the code diff, figure and reproduction command. Check missing columns, unit conversions and the k = 0 limit against the supplied reference. Report checks actually run, and label anything you could not verify.

Acceptance: rerun the command, inspect units and independently check the reference value. A plausible-looking plot is not validation.

OPTIONAL · CT1 · CLAUDE + OVERLEAF Skip tutorial →

Choose a handoff you can review

Start from a clean copy of the Overleaf source and keep the original recoverable.

Source ZIP · broadest access

Overleaf: File → Download as source (.zip).
Extract into a new local working folder.
Work on that copy; upload it as a new Overleaf project.

The source ZIP does not include the compiled PDF.

Git · when enabled for the project

Overleaf Git integration is a premium feature.
Clone or pull the project, then review locally.
Direct Overleaf Git has one branch: usually main for new clones.

Confirm remote + branch. Git pushes can displace active comments or Track Changes.

Overleaf’s current download workflow is File → Download as source (.zip). The archive contains project source files, not the separately generated PDF. Extract the files into a new working directory so the downloaded copy is distinct from the live Overleaf project.

Overleaf Git integration is a premium feature for Overleaf Cloud and is also available on Overleaf Server Pro. Its Git bridge supports a single linear history and one branch. Current documentation says new clones use main; older clones may still use master. Confirm the actual branch before pulling or pushing. Do not assume the Overleaf remote behaves like a normal multi-branch GitHub repository. Overleaf also warns that Git pushes can displace Track Changes and comments, so coordinate with coauthors before using Git on an actively annotated project. GitHub synchronization is a separate feature with its own access requirements.

OPTIONAL · CT2 · CLAUDE + OVERLEAF Skip tutorial →

Install the interface you want to use

VS Code extension for an editor panel, or the CLI for a terminal session.

VS Code extension

Open Extensions: Cmd/Ctrl + Shift + X.
Find Claude Code and install it.
Open your manuscript folder, then the Claude panel.
Sign in when prompted.

Requires VS Code 1.94 or later.

CLI · native install

# macOS / Linux / WSL
curl -fsSL https://claude.ai/install.sh | bash

# Windows PowerShell
irm https://claude.ai/install.ps1 | iex

Open a new terminal after installation; sign in on first run.

The extension is installed from VS Code’s Extensions view. On macOS the shortcut is Cmd+Shift+X; on Windows and Linux it is Ctrl+Shift+X. Open the paper’s folder as the VS Code workspace, then open Claude Code from its panel and complete the sign-in flow. The extension requires VS Code 1.94.0 or later; the supported account options are listed in Anthropic’s documentation.

The code block shows Anthropic’s native installer commands for macOS/Linux/WSL and Windows PowerShell. Windows CMD uses a different command; see the quickstart before pasting a PowerShell command into CMD. After install, open a fresh terminal, check the version if needed, move into the manuscript folder, and run claude. Do not run a long-lived Claude session from a parent folder that contains unrelated files.

OPTIONAL · CT3 · CLAUDE + OVERLEAF Skip tutorial →

Initialize context from inside the session

Terminal commands start the tool; slash commands control a running conversation.

Terminal · before the session

cd /path/to/paper
claude

Open the prompt in the manuscript’s project directory.

Claude Code · after it opens

/init

Review the proposed starter CLAUDE.md before you rely on it.

Use /init in the session — not claude init in the shell.

OPTIONAL · CT4 · CLAUDE + OVERLEAF Skip tutorial →

Keep the project rulebook short and testable

Put stable manuscript conventions and boundaries in the project’s CLAUDE.md.

Starter example

# Paper workflow
- Main file: main.tex
- Bibliography: refs.bib
- Preserve claims, values, units, and cite keys.
- Never invent data or references.
- Edit only the requested section.
- Ask before changing figures or .bib files.

Replace the example filenames with this project’s real files.

Write rules that can be checked

Name the main TeX file and bibliography.
State what may change and what must stay fixed.
Ask for a diff and a compile check.
Keep personal preferences out of shared project rules.

A project-level CLAUDE.md is plain Markdown loaded as context for sessions in that project. Use it for information the assistant should repeatedly know: the manuscript’s main file, bibliography file, compilation convention, and scope limits. Replace main.tex and refs.bib with the actual filenames before sharing the file.

Good boundaries for a scientific manuscript include preserving equations, units, numerical values, citation keys, and the meaning of claims; not inventing data or references; editing only named sections; and asking before changing figures or bibliography records. Ask Claude to report the files changed, the compile result, and unresolved issues. Keep the file concise: Anthropic treats CLAUDE.md as context, not an enforced permission boundary. Use product permission settings for actual tool approvals.

Claude Code: How Claude remembers your project

OPTIONAL · CT5 · CLAUDE + OVERLEAF Skip tutorial →

Make one repeatable edit into a local skill

A project skill is a Markdown workflow Claude can discover or you can invoke by name.

Create this file

.claude/skills/revise-paragraph/SKILL.md

---
name: revise-paragraph
description: Revise one selected paragraph for clarity while preserving its meaning.
---
1. Edit only the named paragraph.
2. Preserve claims, numbers, units, and cite keys.
3. Add no sources; show a diff first.

Discover and invoke

/reload-skills
/skills
/revise-paragraph

Use the skill list to confirm what this project exposes.

For external discovery, browse skills.sh ↗; treat entries as third-party candidates.

Create the folder and save the complete example as .claude/skills/revise-paragraph/SKILL.md at the project root. A skill file uses YAML frontmatter between --- markers followed by Markdown instructions. The folder name or frontmatter name determines the command; the description helps Claude decide when the skill is relevant.

After saving the file, run /reload-skills inside Claude Code to rescan skill directories, then use /skills to list what is available. Invoke this one explicitly with /revise-paragraph, then provide the paragraph to work on. If it is missing, check that the file is under the active project’s .claude/skills/ directory and that the session is running from that project. The local skill path is documented by Anthropic; skills.sh is an external discovery directory, not Anthropic product documentation. Inspect any external skill’s source, scripts, dependencies, and license before adopting it.

OPTIONAL · CT6 · CLAUDE + OVERLEAF Skip tutorial →

Start with one paragraph and a clear stop point

Ask for an evidence-bound edit that is easy to compare with the source.

“Revise only the Results paragraph in sections/results.tex. Preserve every scientific claim, number, unit, and citation key. Add no sources or data. Show a diff and wait for my approval before applying the edit.”

Replace the path with a real section. You remain responsible for the scientific interpretation.

OPTIONAL · CT7 · CLAUDE + OVERLEAF Skip tutorial →

Review the diff, then compile the manuscript

A clean build checks syntax and references; it does not validate the science.

Inspect what changed

/diff

Confirm only the intended files changed.
Check equations, values, units, and citation keys.

Compile and inspect the PDF

# local TeX installation
latexmk -pdf main.tex

Or click Recompile in Overleaf. Check the PDF and build log.

/diff is a Claude Code session command for reviewing changes in the working tree. Inspect the diff in Claude Code or VS Code before accepting it. Verify the changed paths and compare scientific quantities, equations, units, citation keys, and statements against the sources already in the project.

If a local TeX installation is available and the manuscript’s main file is main.tex, Overleaf’s LaTeX guidance documents latexmk -pdf main.tex as a one-command build for a document with references. Replace the filename with the actual main file. Otherwise, compile in Overleaf with Recompile. Overleaf uses latexmk to run the extra passes required for citations and cross-references; still inspect unresolved-citation messages and confirm the rendered bibliography and in-text citations in the PDF. A successful compile does not establish that a scientific claim is correct.

OPTIONAL · CT8 · CLAUDE + OVERLEAF Skip tutorial →

Return only a reviewed copy

Keep ZIP handoffs separate; send Git changes to the confirmed target.

ZIP → new Overleaf project

Package the reviewed source files.
New Project → Upload Project.
Recompile and inspect the new project.

The original project remains a separate project.

Git → confirmed remote + branch

Clean status; pull confirmed main/master.
Review diff, compile, commit named files only.
Fetch, sync, review again; push without force.

New clones use main; older ones may use master. Full command sequence →

If the push is rejected or conflicts appear: stop, sync, review again. Never force-push a manuscript handoff.

Complete tutorial, copyable examples & official sources →

ZIP path: include the manuscript’s source files in the archive, then choose New Project → Upload Project in Overleaf. This creates a new Overleaf project from the ZIP. Recompile there and compare the generated PDF with the reviewed local build. The source ZIP from Overleaf does not include its compiled PDF, so download the original PDF separately if you need a comparison baseline.

Git path: the companion handout contains a complete fresh-clone sequence: confirm a clean state, pull the current Overleaf target, create a local paper-review branch, edit, diff, compile, stage only named files, commit, fetch and rebase on the latest target, re-review, then push paper-review:main (or :master for an older clone). For the direct Overleaf Git bridge, use the actual remote name configured for the project. Current Overleaf documentation says new clones use main; older clones may still use master. The Git bridge supports one remote branch and documents mapping a local review branch to that target. Overleaf warns that pushes can displace active Track Changes and comments; coordinate with coauthors first. Do not use a force push. If the push is rejected or a rebase conflict appears, stop, synchronize with the latest project version, resolve and review the diff again, then compile before retrying. For GitHub or another host, use the project’s actual protected/approved target branch and its normal review process; do not assume this is the Overleaf Git bridge’s branch behavior.

AI · PRACTICAL RESEARCH WORKFLOWA9

Prompts and attachments: make the output testable

One concrete manuscript request is more useful than “improve my paper”.

Attach an evidence pack

manuscript.tex — editable source
references.bib + PDFs — citation evidence
results.csv + units — numerical truth
figures + captions — visual context

Ask the agent to list files it could actually read.

Request a bounded change

Revise the Introduction only. Preserve scientific claims, uncertainty and citation keys. Return a diff and a claim–source table. Mark unsupported claims; do not invent references.

AI · PRACTICAL EXAMPLESA9b

Bad prompt / Good prompt

The same task — a vague request versus clear instructions you can verify.

Bad prompt

Improve my paper.

No target section, source of truth, permitted changes or review format.

Attach

Introduction.tex
references.bib
the cited PDFs

Good prompt

Revise only Introduction.tex for clarity; keep it under 700 words. Preserve scientific claims, uncertainty, equations and citation keys.

Use the attached papers as evidence. Flag unsupported claims with the source passage or “source missing”.

Return a diff and a claim–source table. List unreadable files. Check that every citation key exists in references.bib.

OPTIONAL · AX9 · SUPPORTING DETAILSkip extras →

Prompt: literature review with traceable evidence

Attach your question, inclusion criteria, seed papers and search cutoff.

Find studies comparing finite-difference and finite-element methods for spin-wave dispersion in confined waveguides. Search through [date]. Include primary papers with an explicit geometry, material parameters and a numerical or experimental reference. Exclude papers that only mention the methods.

Return a table: citation / DOI, geometry, method, validation, limitations and supporting page or passage. Separate full texts you inspected from abstract-only records. Do not invent missing values. End with unresolved questions and a reproducible search log.

Acceptance: verify DOI/title matches and inspect the source passages for three key claims. A search is not necessarily exhaustive.

AI · PRACTICAL RESEARCH WORKFLOWA10

Count the whole task — not just the prompt

API tokens, subscription allowance and benchmark cost are different measures.

Illustrative API calculation

20,000 input + 5,000 billed output tokens
No cache, tools or retries

OpenAI model	Estimated text cost
Luna MAX	$0.0045
Terra	$0.045
Sol	$0.09
Astra	$0.45

What increases actual usage

Repeated context and large tool outputs
Reasoning tokens and long answers
Retries, delegation and review
Search, execution and storage charges

Measure: cost per accepted result, plus your review time.

OPTIONAL · AX11 · SUPPORTING DETAILSkip extras →

Prompt: challenge a result before submission

Attach the derivation, assumptions, figure and numerical settings.

Review the attached derivation and dispersion comparison as a critical colleague. Separate algebra, dimensions, physical assumptions and numerical error. Check boundary conditions, k = 0 and the stated limiting case.

Return findings ordered by scientific impact, each with a source location, explanation and a concrete test. Do not silently fix the derivation. Distinguish a confirmed error from a question requiring more evidence. Suggest the single most informative next check.

Acceptance: resolve each material finding with an independent calculation or source; agreement between models is not sufficient.

AI · PRACTICAL RESEARCH WORKFLOWA11

Quality and cost: three complementary comparisons

Read each score together with its task, reasoning setting and cost definition.

Selected Artificial Analysis models: benchmark score versus benchmark cost — **Artificial Analysis** · selected models and reasoning settings

Next: ARC-AGI-1 →
Visual reasoning · score vs task cost

Then: subscription value →
Index score vs estimated subscription cost

Different tests and cost definitions — compare within each chart. Full-size views follow immediately.

AI · QUALITY, REASONING & COSTA11b

Reasoning effort: compare the gain with the cost

ARC-AGI-1 · abstract visual reasoning · score versus cost per task

Read the axes

Up = higher score; left = lower cost. The cost axis is logarithmic.

Follow a model’s curve

More reasoning can improve the score, with diminishing gains near the top.

Choose for your task

Use this benchmark to shortlist settings, then test your own research tasks.

Explain ARC-AGI-1 as abstract visual reasoning using grid transformations; it is not a comprehensive measure of scientific reasoning, manuscript quality or coding. The source leaderboard explains that connected reasoning-system points represent one model at different reasoning levels and that gains typically flatten. Each tenfold x-axis interval has equal visual width; it is cost per benchmark task, not per million tokens. In the supplied image several curves approach the top, so this benchmark has limited headroom for distinguishing those configurations. Prefer points towards the upper left when comparing the same test and evaluation protocol, then consider latency and your own acceptance tests. Do not claim that the most expensive setting necessarily wins or that every curve is monotonic. Do not infer exact model rankings or unlabelled values from tiny points. The captured ARC-AGI-1 figure is preserved as submitted; the live leaderboard currently opens on ARC-AGI-3, a different benchmark. No claim is made that this screenshot is the latest ranking.

AI · QUALITY, REASONING & COSTA11c

Subscription value: score is only half the picture

A second cost model · the supplied mypriors.com chart compares index score and subscription cost per task

Look toward the upper left

Higher score at lower estimated cost is attractive within this comparison.

Compare effort settings

Sol high, Astra max and Claude variants occupy different score–cost positions.

Keep the cost models separate

Subscription cost per task is not API token pricing or a guaranteed marginal fee.

Read the vertical axis as the chart’s selected index score, not a percentage of ARC-AGI tasks solved. The horizontal axis is labelled subscription cost per task in dollars on a logarithmic scale. A logarithmic axis cannot plot zero; the dashed free column is a special category, not a numeric logarithmic coordinate. The black connected envelope highlights favourable score-cost trade-offs among the displayed points, but does not establish universal model superiority. The provider site supports configurable benchmark mixes; the exact weights, subscription allocation, task denominator, pricing assumptions and meaning of fallback for this particular screenshot were not supplied and could not be recovered from the rendered builder page. Retain fallback labels as part of the tested configuration, not as pure single-model results. Subscription allocation depends on plan and use, while API charges depend on billable input/output/cache tokens. Do not use this chart to calculate the author’s API equivalent on A4b. Use it as a shortlist for a controlled comparison on representative manuscripts, analysis scripts and numerical tasks. Treat the image as a supplied snapshot, not a verified current universal ranking.

OPTIONAL · AX4 · SUPPORTING DETAILSkip extras →

Professional agent tasks: score versus cost

User-supplied chart · OpenAI vendor-reported benchmark · compare within this test.

Read the model, reasoning setting, task suite and cost definition together. This is not an independent scientific-research ranking.

OPTIONAL · AX5 · SUPPORTING DETAILSkip extras →

Terminal work: a different cost–quality curve

User-supplied chart · OpenAI vendor-reported benchmark · compare within this test.

Read the model, reasoning setting, task suite and cost definition together. This is not an independent scientific-research ranking.

OPTIONAL · AX6 · SUPPORTING DETAILSkip extras →

An independent coding-agent comparison

Vals AI · Terminal-Bench 4.0 · 26 October 2026 · Mini-SWE-agent

Model	Score	Reported cost / test
Claude Opus 5.5	65.15%	$13.20
Claude Sonnet 5.5	64.14%	$16.51
Claude Fable 5.1	58.08%	$17.18
Gemini 4 Argon	57.58%	$17.64
Claude Haiku 5.5	36.36%	$3.12

Same evaluator and harness, selected rows. Different samples and settings from the OpenAI chart. Measure your own accepted outputs.

AI · PRACTICAL RESEARCH WORKFLOWA12

Six habits that save tokens without losing rigor

Start economical; spend stronger reasoning where a failed answer is costly.

01 · Scope the context

Attach the relevant sections; search before reading entire folders.

02 · Bound the output

Ask for a patch, table or decision — not repeated full reports.

03 · Keep a handoff

Summarise accepted facts and open issues before starting a new task.

04 · Route by difficulty

Strong model for the plan; smaller model for explicit repetitive steps.

05 · Reuse evidence

Do not repeat unchanged searches or tests. Recheck after relevant changes.

06 · Measure the loop

Include retries, tool calls, review and corrections in the total cost.

FULLMAG · THE PROJECTF0

FullMag

A research platform for finite-difference and finite-element micromagnetics.

One study model.
Multiple numerical engines.

From magnetic textures and dynamics to eigenmodes and dispersion.

Mateusz Zelent · Philipp Pirro

RPTU Kaiserslautern-Landau, Germany
Department of Physics & OPTIMAS

Mateusz Gołebiewski · Uladzislau Makartsou

Adam Mickiewicz University, Poznań, Poland
Faculty of Physics and Astronomy

Funded by
the European Union

Mateusz Zelent acknowledges funding from the European Union's Horizon Europe programme under HORIZON-MSCA-2024-PF-01, Marie Skłodowska-Curie Grant Agreement No. 101208951 (CNMA).

Project coordination: Mateusz Zelent · RPTU Kaiserslautern-Landau

FULLMAG · THE PROJECTF0b

Why build FullMag?

Start with the research question — then choose the geometry, physics and numerical method.

01

Explore new physics

Add interactions and study magnetic textures, coupled structures and spin-wave modes.

02

Choose the right method

Use regular grids for films and multilayers, or conforming meshes for curved and complex geometries.

03

Keep the experiment connected

Carry one study from model setup through execution to field maps, spectra and reproducible results.

Can we turn a new micromagnetic idea into an inspectable numerical experiment — within one workflow?

The goal: a platform we can extend, inspect and compare with established solvers.

FULLMAG · THE PROJECTF0c

Started in April — developed with AI

Mateusz Zelent began FullMag in April, using AI assistance from the start.

APRIL

An idea becomes a project

Start building a micromagnetics platform around real research needs.

ITERATION

AI-assisted development

Connect scientific reasoning, implementation, debugging and documentation.

TODAY

A shared research framework

FDM + FEM, Python + Control Room, dynamics + spectral analysis.

AI tools used in the development workflow

OpenAI

ChatGPT · Codex

Anthropic

AI model provider

Claude

Chat · coding assistance

Gemini

Google AI

AI assists the work; the team owns the physics, implementation choices and validation.

OPTIONAL · G11 · SUPPORTING DETAILSkip extras →

Docs public overview

The public landing page assigns browser workflow, solver/physics/runtime facts, authoring API, and validation to separate documentation families; its flow diagram distinguishes requested intent from resolved execution.

FULLMAG · METHODS & IMPLEMENTATIONF1

Two discretizations, two different strengths

FDM = finite differences · FEM = finite elements. Both approximate the same continuum magnet.

Question	FDM · regular cells	FEM · unstructured mesh
How is the sample represented?	Magnetization on cuboid cells; surfaces cut by an active mask	Magnetization expanded in nodal basis functions on mesh elements
How are derivatives computed?	Neighbour stencils and face differences	Integrals of basis functions and their gradients: the weak form
How is demagnetization computed?	Nonlocal tensor convolution, accelerated by FFT	Scalar-potential PDE; finite airbox or FEM–BEM open-boundary treatment
Where is it attractive?	Extended films, regular waveguides and repeated GPU workloads	Curved boundaries, local refinement and complex 3D geometry
What must converge?	Cell size, boundary representation and time step	Mesh size/order, exterior treatment and linear-solver tolerance

Choose the discretization from the geometry and required observable; compare accuracy before speed.

FULLMAG · METHODS & IMPLEMENTATIONF2

Why Fullmag alongside MuMax3 and COMSOL?

A research platform joining micromagnetics, solver development and inspectable workflows.

	Fullmag	MuMax3	COMSOL
Core approach	One model; FDM and FEM execution paths	Specialized GPU finite-difference micromagnetics	General finite-element multiphysics; micromagnetics requires a suitable formulation/module
Research advantage	Own operators, new interactions, mesh/grid choice and analysis products	Established reference for films, dynamics and micromagnetic benchmarks	Geometry, coupled PDEs and multiphysics modelling tools
Workflow	Python + Control Room → shared model → plan → saved results	Scripted simulation and output analysis	Model/study setup with configurable solvers
Trade-off	We maintain physics, cross-language interfaces and qualification	Grid representation must resolve curved boundaries	Licensing, model setup and solver choices; no universal performance comparison

Our motivation: change the physics and numerical method without rebuilding the entire research workflow.

FULLMAG · METHODS & IMPLEMENTATIONF3

FullMag architecture: one study, two numerical engines

Who describes the experiment, who computes the physics, and who turns results into insight?

FDM/FEM select the discretization; LLG, relaxation and eigensolve select the mathematical task.

55 seconds · 13:49–14:44

A preconditioner makes an iterative solve converge more efficiently; it does not alter the physical PDE. CEED is an operator execution layer rather than a complete mesh manager or global linear solver. Bare CMake defaults the MFEM stack off; managed profiles select it.

Implementation audit of current main source; no fresh solve or scientific qualification was run for this presentation. Screenshot evidence has its own source date/commit.

Library details retained from the previous F3:

Library	General role	What Fullmag uses / configures
MFEM	C++ finite-element discretization library	Mesh and H1 spaces; weak forms, stiffness/mass operators, potential and field recovery
HYPRE	Sparse iterative linear algebra	PCG/GMRES and BoomerAMG for demagnetizing-potential systems; device-capable HYPRE in strict GPU routes
libCEED	Portable, matrix-free/partially assembled element operators	Built with MFEM and selected by the managed GPU profile; it is not an alternative demag equation
CUDA + cuFFT	Device kernels and Fourier transforms	Fullmag FDM stencils, FFT convolution and stepping; FEM RHS/recovery and supported demag kernels

Pinned build profile

MFEM 4.10 · HYPRE 3.1.0 · libCEED 0.12.0. Managed CPU and GPU profiles build different dependency configurations.

Important libCEED distinction

MFEM is configured with CEED support, but Fullmag’s availability record currently reports libceed_used_hot_path = 0. Do not claim that every operator runs through CEED.

Read the diagram from top to bottom. Python and Control Room are authoring routes, not separate scientific engines. ProblemIR and ExecutionPlan connect the shared description to the chosen runtime. FDM FFT spectra are computed from time trajectories, whereas FEM eigensolve linearizes around equilibrium and solves a separate spectral problem; driven response is a further distinct linearized solve. The solver boxes are conceptual families, not a claim that every interaction/device/precision combination is admitted. Rust records artifacts and provenance and also provides the CPU FDM implementation. MFEM supplies FE spaces and operators, HYPRE sparse solves/preconditioning. PETSc/SLEPc support selected spectral routes; CUDA/cuFFT accelerate supported execution. libCEED support is configured, while the recorded availability flag libceed_used_hot_path=0 prevents a claim of universal CEED execution. Fine-grained gates and qualification remain on E4 and the detailed method slides. Diagram synthesizes the already-audited F3, F4 and E4 evidence; no new solver run was performed.

OPTIONAL · G14 · SUPPORTING DETAILSkip extras →

Docs ui architecture webgl

The architecture page documents the Three.js/React Three Fiber viewport, vector rendering, resource invalidation, and context recovery boundaries.

FULLMAG · METHODS & IMPLEMENTATIONF4

What runs in Python, Rust, C++ and CUDA?

A shared scientific description feeds separate numerical implementations.

Layer	Technology	Responsibility in Fullmag
Author the experiment	Python DSL · NumPy · optional Gmsh	Geometry, materials, interactions, staged studies and mesh preparation
Validate and plan	Rust · typed ProblemIR and ExecutionPlan	Resolve FEM/FDM, CPU/GPU, precision and supported combinations
Execute and record	Rust runner · C ABI into native backends	Run lifecycle, output streams, field data and provenance; Rust also owns the FDM CPU reference engine
Compute fields and evolve m	C++/MFEM and CUDA	FEM operators/solves; FDM CUDA stencils, cuFFT demag and integrators; FEM device kernels where supported
Inspect the science	TypeScript · React/Next · Three.js · ECharts	3D geometry/fields, dispersion curves, result tree and controls; Axum/Tokio provide the Rust API

Python describes the study; Rust coordinates it; numerical kernels execute the selected realization.

FULLMAG · METHODS & IMPLEMENTATIONF5

Micromagnetics: energy determines the field; LLG determines motion

The state is the unit magnetization m(r,t); material parameters set the competing length and time scales.

M = M_{s} m, | m | = 1, δ E = - μ_{0} \int_{Ω_{m}} M_{s} H_{e f f} \cdot δ m d V

\partial_{t} m = - \frac{γ_{0}}{1 + α^{2}} [m \times H_{e f f} + α m \times (m \times H_{e f f})] + T_{e x p l i c i t}

What the terms do

Precession: rotation around the effective field.
Damping: energy loss towards a stationary state.
Direct torque: current-driven rates added with the correct Gilbert conversion.

Symbols and units

Mₛ, H: A/m; E: J; m, α: dimensionless.
γ₀ = μ₀|γ|: m/(A·s); T: s⁻¹.
Input B in tesla must be converted consistently; B and H are not interchangeable.

The same LLG model can be discretized with FDM or FEM; that choice changes the numerical operator.

FULLMAG · METHODS & IMPLEMENTATIONF6

Exchange illustrates what “grid” and “weak form” mean

Exchange penalizes rapid spatial changes; the discretization turns this energy into a field.

E_{e x} = \int_{Ω_{m}} A | \nabla m |^{2} d V, H_{e x} = \frac{2}{μ_{0} M_{s}} \nabla \cdot (A \nabla m)

FDM: neighbours and cell faces

E_{h} = \sum_{f = i | j} A_{f} \frac{S_{f}}{d_{f}} | m_{j} - m_{i} |^{2}

Each face couples adjacent cell values.
Face area S and distance d encode geometry.
Material interfaces require a defined face coefficient A_f.

FEM: basis functions and integrals

K_{i j} = \int A \nabla ϕ_{i} \cdot \nabla ϕ_{j} d V, M_{s}^{(h)} h = - \frac{2}{μ_{0}} K m

φᵢ: basis function; K: exchange stiffness matrix.
Weighted mass projection converts the energy residual into a field.
Consistent and lumped projection are distinct numerical choices.

A [J/m] controls stiffness; resolving the exchange length matters in both discretizations.

FULLMAG · METHODS & IMPLEMENTATIONF7

Conservative interactions: more than exchange and demag

These terms change the energy and effective field; their numerical realization depends on the solver.

Interaction	Physical effect	Current implementation scope
Exchange	Penalizes neighbouring spins pointing differently	FDM CPU/CUDA; FEM weak operator; multilayer intralayer exchange
Uniaxial + cubic anisotropy	Preferred axes / crystal directions	FDM and FEM paths; material-field and topology restrictions are lane-specific
Zeeman + demagnetization	Applied field + nonlocal self-field	Uniform Zeeman; FDM FFT / multilayer convolution; FEM Poisson / FK
Interfacial DMI	Chiral gradients linked to an interface normal	Single-grid FDM; FEM source paths; multilayer global +z interfacial model
Bulk DMI	Chiral volume gradients, m·curl(m)	Single-grid FDM and FEM source paths; public multilayer rejects bulk DMI
Rotated interfacial DMI	Fixed chiral operator for in-plane-background textures	Implemented FDM/FEM CPU/GPU within explicit topology/boundary gates
Prescribed-strain magnetoelastic field	Strain modifies magnetic energy	Internal reference operators; not a qualified coupled mechanics solver

“Implemented” describes an operator path; the next question is which geometry, device and workload is admitted.

55 seconds · 17:12–18:07

This replaces the earlier incomplete interaction slide. Conventional DMI FEM source exists but broad time-domain promotion is narrower; rotated DMI has its own explicit operator and restrictions. RKKY/interlayer exchange is not included as executable: authoring objects exist but current planners reject it.

Implementation audit of current main source; no fresh solve or scientific qualification was run for this presentation. Screenshot evidence has its own source date/commit.

FULLMAG · METHODS & IMPLEMENTATIONF8

Torques, noise and current-generated fields

A direct torque is a rate in s⁻¹; an effective field is in A/m. They enter LLG differently.

Mechanism	What it represents	Current scope / important boundary
Prescribed SOT	Field-like and damping-like spin-orbit torques	CPU reference and bounded CUDA/FEM paths; prescribed SOT is not a solved spin-Hall transport model
Slonczewski STT	Spin-polarized current acting on a thin magnetic layer	Bounded fullmag.v2 local target-masked routes; interface-flux variant unsupported
Zhang–Li STT	Current couples to magnetization gradients	Separate formula/version contracts: MuMax3-v1 SP5 evidence does not validate Fullmag-v1 or every lane
Thermal Brown noise	Stochastic field determined by temperature, damping, volume and time step	Single-grid FDM CPU/GPU; FEM and public multilayer thermal are not executable here
Oersted field / current transport	Magnetic field generated by a declared or computed current	Named cylinder and prescribed-density paths; separate antenna field-solve workflow
RKKY / interlayer exchange	Exchange coupling across layers	Python/IR objects exist; current planner rejects executable coupling

Multilayer magnetostatic coupling is implemented separately from interlayer exchange and spin torques.

55 seconds · 18:07–19:02

Public multilayer currently rejects spin-torque/current bindings, thermal, Oersted and magnetoelastic. Current FDM CUDA OerstedCylinder planner admits +z and absent/constant time dependence only; the capability summary lists broader envelopes and is inconsistent. Some torque paths are reference-executable or bounded production workloads; do not call all four solver/device lanes equivalent. Current source evidence is in the accompanying refreshed audit.

Implementation audit of current main source; no fresh solve or scientific qualification was run for this presentation. Screenshot evidence has its own source date/commit.

FULLMAG · METHODS & IMPLEMENTATIONF9

Rotated interfacial DMI is implemented

A distinct chiral energy for in-plane-background textures such as bimerons — not merely a rotated easy axis.

E_{r D M I} = \int_{Ω_{m}} D [m_{z} \partial_{x} m_{x} - m_{x} \partial_{x} m_{z} + m_{x} \partial_{y} m_{y} - m_{y} \partial_{y} m_{x}] d V

H_{r D M I} = \frac{2 D}{μ_{0} M_{s}} {(\partial_{x} m_{z} - \partial_{y} m_{y}, \partial_{y} m_{x}, - \partial_{x} m_{x})}^{T}

Meaning and authoring

D [J/m²] sets strength and chirality; signed D is preserved.
RotatedInterfacialDMI(D) selects this fixed operator.
anisU sets anisotropy separately; it does not rotate DMI.

Implemented scope

FDM: CPU and CUDA, including bounded multilayer routes.
FEM CPU: P1 tet4 / prism6 / pyramid5; GPU: P1 tet4.
Nonzero FEM periodic and eigen/frequency combinations are rejected; current Göbel bimeron qualification is NOT VERIFIED.

Nonzero rDMI at open/masked boundaries requires Exchange with positive local A; the natural boundary includes both terms.

55 seconds · 19:02–19:57

The operator is D(L_zx^x+L_xy^y), equivalent to D21=D32=D in the documented tensor convention. Do not claim an arbitrary runtime rotation angle. Source implementation does not establish that every bimeron setup is qualified. Cannot combine nonzero rotated DMI with conventional/material interfacial or bulk DMI. D=0 preserves intent as a no-op.

Implementation audit of current main source; no fresh solve or scientific qualification was run for this presentation. Screenshot evidence has its own source date/commit.

FULLMAG · ROTATED DMI RESULTSF9b

Rotated DMI: simulating bimerons

From magnetic textures to relaxation, a held state and energy diagnostics.

Stabilization report · initial → relaxed → held

Frozen-profile energy · diagnostic scan

Bimerons in Fullmag: initialize the texture → relax → hold → inspect the state and energy.

55 seconds · 19:57–20:52

Author-supplied bimeron results. The stabilization report labels rotated interfacial DMI, strict CUDA FDM FP64 and VERIFIED 15/15 gates; it shows initial, relaxed (20 ps) and held (120 ps) states. These are claims recorded in the supplied report, not a fresh rerun for this deck. The separate frozen-profile energy plot labels its open circles diagnostic / failed and must not be presented as a qualified equilibrium-radius curve. The workspace capture shows an intermediate frozen-spin run. Do not infer identical run settings across all five images. Click each original image to inspect labels at full resolution.

Implementation audit of current main source; no fresh solve or scientific qualification was run for this presentation. Screenshot evidence has its own source date/commit.

Recorded image provenance: user-bimeron-manifest.json

FULLMAG · METHODS & IMPLEMENTATIONF10

FDM demagnetization: a nonlocal tensor convolution

Every magnetic cell contributes to the field in every other cell; FFTs accelerate this long-range sum.

H_{d, i} = - \sum_{j} N_{i j} M_{j}, {\hat{H}}_{d} = - \hat{N} \hat{M}

Operator and algorithm

Newell tensor N: volume-averaged cuboid interaction, including the self term.
Pad for linear convolution → forward FFT of M → tensor–vector multiplication → inverse FFT → crop.
The kernel is reused while geometry and the relevant discretization stay fixed.

What runs where

FDM CPU: Rust reference engine and rustfft.
FDM GPU: CUDA buffers/kernels and cuFFT; FP32/FP64 are distinct paths.
Open and periodic models require different kernel/boundary conventions.

N is dimensionless in this convention; M and H are A/m. FFT speed does not remove spatial-discretization error.

FULLMAG · METHODS & IMPLEMENTATIONF11

Multilayer convolution: couple layers without filling the gaps

Each destination layer receives a field from every source layer through its own oriented pair kernel.

{\hat{H}}_{i} (k) = - \sum_{j = 1}^{L} {\hat{N}}_{i j} (k) {\hat{M}}_{j} (k)

Stage	Operation	Why it is needed
1 · Native magnetization	Mⱼ = Mₛ,ⱼmⱼ → common scratch representation	Preserve the physical magnetic moment when layouts differ
2 · Transform	Pad and FFT each source layer	Reuse a source spectrum across its destination layers
3 · Couple ordered pairs	Accumulate Nᵢⱼ Mⱼ for every j → i	Kernel includes source/destination thickness and physical offset
4 · Recover fields	Inverse FFT, crop and pull back to native layer	Return Hᵢ on the actual magnetic cells, not on an artificial spacer grid

There are L² ordered source–destination pairs; equal kernel descriptors can share cached kernels.

FULLMAG · METHODS & IMPLEMENTATIONF12

Multilayer: two geometrical models and a concrete work programme

The CPU descriptor path and CUDA path have different current coverage.

	two_d_stack	three_d
Layer representation	One native z cell per layer, unless explicitly averaged	Several z cells per volumetric layer
Transform	2D FFT in x/y; scratch z extent = 1	3D pair convolution over signed x/y/z lags
Thickness and gaps	Independent thickness and centre offsets; shared in-plane cell pitches required	Layer origins retained; translational cell pitches must match for FFT pairs
Meaning	No fictitious magnetic cells in a spacer	Resolves magnetization variation across layer thickness

What exists

CPU FP64 ordered-pair descriptors, Newell kernels and explicit transfers.
CUDA FP64/FP32 source; bounded 3D identity contract, with the whole step still assisted.

What is still being completed

Canonical CPU/CUDA pair-kernel parity and composed unequal-thickness tests.
CUDA XY offsets / heterogeneous thickness / transfers; fresh device evidence.
Current qualification matrix does not qualify the full CUDA multilayer route.

This is active numerical development, with working components and explicit remaining end-to-end gaps.

55 seconds · 22:33–23:28

CPU local oracle/identity results are bounded, not production qualification. CUDA 2D-self currently uses first-layer cell parameters and a z-only shifted kernel rather than the CPU descriptor pair operator. A general multilayer claim would hide this important gap. Positive-volume overlap is rejected; uncertain imported/CSG overlap fails closed.

Implementation audit of current main source; no fresh solve or scientific qualification was run for this presentation. Screenshot evidence has its own source date/commit.

FULLMAG · METHODS & IMPLEMENTATIONF13

FEM demagnetization I: Poisson on magnet + air

Magnetization creates a potential; the surrounding air is part of the computational mesh.

H_{d} = - \nabla u, \nabla \cdot (H_{d} + M) = 0, Δ u = \nabla \cdot M

\int_{Ω_{m} \cup Ω_{a}} \nabla u \cdot \nabla v d V = \int_{Ω_{m}} M \cdot \nabla v d V

Numerical sequence

Assemble the source from M = Mₛm in magnetic elements only.
Solve the scalar potential system on the shared magnetic + air mesh.
Recover H_d = −∇u and magnetic-domain energy.

Outer boundary is an approximation

Dirichlet: u = 0 at the artificial boundary.
Robin: ∂ₙu + βu = 0, adding a boundary matrix; β has units m⁻¹.
Vary airbox size and mesh independently to test truncation error.

u [A], M and H [A/m]; the finite airbox approximates an unbounded exterior.

55 seconds · 23:28–24:23

For test functions vanishing on the Dirichlet outer boundary the shown weak form has positive ∫M·∇v. Robin adds β∫boundary uv to the left. Fullmag chooses β=c/R* from a configured factor and characteristic mesh extent. Potential u has unit A. A free magnetic surface is not the artificial outer air boundary.

Implementation audit of current main source; no fresh solve or scientific qualification was run for this presentation. Screenshot evidence has its own source date/commit.

OPTIONAL · G8 · SUPPORTING DETAILSkip extras →

Airbox setup and explicit availability feedback

Airbox setup for the mixed-P1 model, with a visible warning that vector quantity m is unavailable and the runtime lane is unresolved.

FULLMAG · METHODS & IMPLEMENTATIONF14

FEM demagnetization II: Fredkin–Koehler

FK combines volume finite elements with a boundary-element description of the exterior.

u = u_{1} + u_{2}, H_{d} = - \nabla (u_{1} + u_{2})

Step	Mathematical task	Fullmag realization
1 · Volume source	Δu₁ = ∇·M; ∂ₙu₁ = M·n	Neumann FEM solve; one gauge constraint per connected component removes the constant null mode
2 · Exterior contribution	u₂\|Γ = B(u₁\|Γ)	Boundary trace → hierarchical BEM operator; near interactions evaluated directly, far interactions compressed
3 · Harmonic correction	Δu₂ = 0 in the magnet	Second FEM solve with boundary values from B; combine and recover −∇u
4 · Energy	E_d = −(μ₀/2)∫ M·H_d dV	Integrate on the magnetic body; the exterior is accounted for by the boundary operator

No airbox mesh: FK represents the open exterior through the magnetic surface Γ.

58 seconds · 24:23–25:21

B here denotes the implemented FK boundary double-layer/jump operator, not magnetic flux density. The notation intentionally encapsulates the code’s normal and jump convention. CPU uses Lindholm near blocks and ACA low-rank far blocks. Dense BEM is a small-mesh oracle with quadratic surface memory; hierarchical compression is the scalable source route. This is FEM–BEM, not a different physical demagnetization law.

Implementation audit of current main source; no fresh solve or scientific qualification was run for this presentation. Screenshot evidence has its own source date/commit.

FULLMAG · METHODS & IMPLEMENTATIONF15

Poisson airbox versus FK: method, cost and current state

Both are implemented; they make different geometric and algorithmic trade-offs.

	Poisson airbox	Fredkin–Koehler FEM–BEM
Unknowns / exterior	Potential on magnet + meshed air	Body potentials + boundary integral; no airbox
Main error controls	Body mesh, air mesh, outer radius and boundary condition	Body/surface mesh, quadrature and far-block compression
CPU path	MFEM assembly + HYPRE scalar solve	Two sparse HYPRE solves + hierarchical surface apply
GPU path in source	CUDA/CSR RHS and recovery + device-HYPRE Poisson	Uploaded BEM hierarchy + CUDA apply + two device-HYPRE solves
Current FK restrictions	Separate periodic Poisson realization exists	3D body-only TET4/P1; nonperiodic; oriented, watertight boundary
Validation status	Scoped evidence; do not generalize a periodic k=0 pass	CPU/GPU implementation present; current physics qualification remains NOT VERIFIED

Time-domain FK support does not imply FK support in eigenmode or frequency-response solvers.

FULLMAG · METHODS & IMPLEMENTATIONF16

Boundary conditions are part of the physical problem

A surface condition for magnetization is different from a boundary condition for the stray-field potential.

Boundary / field	Condition and meaning	Consequence for simulation
Free exchange surface · m	∂ₙm = 0 when exchange alone supplies the surface variation	No imposed surface twist; do not reuse this blindly when DMI is active
Exchange + DMI · m	Natural boundary balances exchange gradient and chiral boundary term	Edge canting is part of the model, including rotated DMI
Fixed / frozen magnetization	Selected magnetization is constrained by the authored condition	Changes admissible dynamics and mode shapes; a pinned mode condition is not an airbox condition
Periodic sample · m	Match paired boundaries; waves acquire a Bloch phase in eigensolve	Boundary pairing and translation must be defined consistently
Open exterior · u	Poisson: finite outer condition. FK: boundary integral for the exterior	Changes demag accuracy and, therefore, equilibrium and frequencies

2 A \partial_{n} m + D (n_{x} m_{z} - n_{y} m_{y}, n_{y} m_{x}, - n_{x} m_{x})^{T} = 0 (exchange + rDMI)

n is the outward unit normal; ∂ₙ is a normal derivative. Boundary choices must accompany every comparison.

OPTIONAL · G9 · SUPPORTING DETAILSkip extras →

Inspect FEM boundary faces · UI fixture

Fullmag’s FEM topology fixture shows the Boundary Faces Inspector and outer-boundary mesh selection; this is UI-fixture evidence only.

FULLMAG · METHODS & IMPLEMENTATIONF17

Time evolution, relaxation and energy minimization

These algorithms answer different questions even when they reuse the same effective-field operators.

Task	Numerical idea	What to inspect
Run dynamics	Integrate LLG with the selected stepper and time/error control	m(t), energy, torque and temporal convergence
Relax with LLG	Suppress precession / follow damping towards a stationary state	Stopping criterion and the final torque; elapsed relaxation time is not a physical transient
Minimize energy	Projected gradient / Barzilai–Borwein choices where admitted	Energy descent, unit-length constraint and termination reason
Prepare an eigenproblem	Accept a stationary state and its effective field	A residual equilibrium torque contaminates the linearization

Implementation choices matter

Fullmag contains explicit stepping and relaxation paths in Rust/CUDA and native FEM, with different admissible integrator/precision combinations. A solver name alone does not specify the run.

Comparisons need the same history

Different random seeds, minimizers or stop rules can choose different metastable states, explaining part of the hysteresis differences shown next.

Save the integrator, time step/tolerance, initial state and actual stopping reason with the result.

OPTIONAL · G3 · SUPPORTING DETAILSkip extras →

YIG waveguide: replay the dynamics

Read-only playback shows the recovered YIG waveguide m_z field at 4 ns (frame 801/801); native run publication had failed after ENOSPC, so this proves playback of recovered data only.

Saved YIG waveguide result; m_z; 4.000 ns; frame 801/801; FDM grid 1000 x 20 x 1; read-only playback.

The Fullmag read-only saved-results viewer displayed recovered YIG-waveguide m_z data at 4 ns. The browser artifact reports 801 frames and verified field payloads.

Limits: The screenshot says numerical data were recovered after ENOSPC but native run publication failed. It proves playback of recovered data only. Browser evidence used requested SwiftShader; it is not hardware-GPU or independent physics qualification. The source browser URL was http://127.0.0.1:3398; that port is offline at the latest collection check.

Source: /mnt/c/Users/Mateusz/Documents/ChatGPT/Fullmag/YIG-playback-FEM-regression-20261028-final.png

Thread: 01a17a79-8ac3-7d60-93a7-e4b184f83113 · Commit: not recorded

Full screenshot provenance

FULLMAG · METHODS & IMPLEMENTATIONF18

Microwave antennas: from current to excitation field

The current distribution and conductor geometry determine the spatial Oersted field applied to the magnet.

3D: conductor geometry and field-vector samples.

Cross-section: |H| colour map and projected field lines.

Stage	What is computed	Why it matters
Conductor / current	Charge-transport solution and conservative current representation	The return path and terminal conditions affect the generated field
Field operator	Direct tetrahedral Biot–Savart quadrature or bounded vector-potential FEM route	Computes the field in the target region; not a full-wave radiation claim
Magnetic excitation	Sample/project the saved field basis; apply the declared drive	A spatially nonuniform drive selects which modes are excited

The two supplied figures visualize saved field data; they are not screenshots of the live Fullmag UI.

55 seconds · 27:45–28:40

Current source has direct quadrature plus a bounded vector-potential reference route. The vector-potential implementation uses ND edge elements, an H1 gauge block, curl-curl and mixed gradient forms, with bounded matrix size. The supplied 3D plot explicitly says inspection data without independent qualification; 321/1145 samples shown. Do not infer RF scattering/S-parameters or full-wave Maxwell from these figures.

Implementation audit of current main source; no fresh solve or scientific qualification was run for this presentation. Screenshot evidence has its own source date/commit.

OPTIONAL · G21 · SUPPORTING DETAILSkip extras →

Microwave antenna: Oersted-field cross-section

Saved-data field plot: colour shows |H|; streamlines show the in-plane direction.

OPTIONAL · G22 · SUPPORTING DETAILSkip extras →

Microwave antenna: conductor and field in 3D

Conductor geometry and sampled Oersted-field vectors around the antenna.

FULLMAG · METHODS & IMPLEMENTATIONF19

Demagnetizing field in the surrounding air

Two views of the airbox: the magnetostatic field extends beyond the magnetic material.

Oblique view · field vectors throughout the airbox

Rotated view · spatial structure around the magnetic bodies

Read the visualization: arrows show field direction; the selected colour mode encodes orientation. Use magnitude colouring and an A/m scale to compare field strength.

Hᵈ = −∇u: inspect the exterior field as well as the magnetization inside the sample.

FULLMAG · METHODS & IMPLEMENTATIONF20

Standard Problem 4 · Fullmag FDM vs MuMax3

Matching snapshots and overlapping magnetization traces

Top: Fullmag FDM GPU single precision. Bottom: MuMax3. Snapshots at 0, 5 and 10 ns.

Agreement at plotting resolution · Fullmag: solid lines; MuMax3: dashed lines.

51 seconds · 29:13–30:04

Use the author-supplied comparison as positive scientific evidence with its scope. Do not invent relative errors or identical input settings where raw inputs were not supplied.

Implementation audit of current main source; no fresh solve or scientific qualification was run for this presentation. Screenshot evidence has its own source date/commit.

The author identifies this run as Standard Problem 4. Exact canonical geometry, applied field, initial state and numerical error norms were not supplied with the figure; no bitwise or zero-error equivalence is asserted.

Details moved from the slide:

What agrees visibly

The spatial snapshots show very similar states.
The component-wise mean magnetization traces nearly overlap.
The transient and subsequent oscillation appear consistent at this plotting resolution.

What this demonstrates

A concrete cross-code dynamical comparison. Numerical error norms and mesh/time-step refinement would turn visual agreement into quantified agreement.

Top: Fullmag FDM GPU single precision. Bottom: MuMax3. Snapshots at 0, 5 and 10 ns.

OPTIONAL · G27 · SUPPORTING DETAILSkip extras →

Fullmag and MuMax3: close agreement in this case

Spatial snapshots and average magnetization traces show close visual agreement at the displayed resolution.

FULLMAG · METHODS & IMPLEMENTATIONF21

Hysteresis: agreement also reveals what to investigate

A field sweep follows metastable states, so preparation and stopping rules can shift switching events.

The supplied comparison uses 128 × 32 × 1 cells, with cell size 4 × 4 × 30 nm.

Read the branches

Large portions of the loop overlap closely.
Down-switch: both report −76 → −77 mT.
Ascending switch: MuMax3 76 → 77 mT; Fullmag 79 → 80 mT.
The initial jump also differs.

Interpretation

The figure reports different random initial states, relaxation/minimization algorithms and stop criteria. These differences need controlled comparison before attributing a shift to one field operator.

A useful benchmark reports differences explicitly and controls the history of the state.

OPTIONAL · G24 · SUPPORTING DETAILSkip extras →

Hysteresis: Fullmag and MuMax3

Branches largely overlap; the figure reports different initial and ascending switching intervals.

FULLMAG · EIGENSOLVE · SEPARATE SOLVER FAMILYE0

Eigensolve: natural oscillations, not time stepping

A separate solver family built around a linearization of the magnetic equations.

Time-domain LLG

Input: initial magnetization and a time-dependent drive.
Output: a trajectory m(r,t), including nonlinear motion.
A spectrum is inferred afterwards from an FFT.

Eigenmode problem

Input: equilibrium m₀, material operators and boundary conditions.
Output: selected eigenfrequencies and complex spatial mode profiles.
Small-amplitude approximation; no time trajectory is integrated.

equilibrium ⟶ linearized operator ⟶ (ω_{n}, δ m_{n})

An animated eigenmode is reconstructed from a complex mode; it is not evidence of a time-domain simulation.

FULLMAG · EIGENSOLVE · SEPARATE SOLVER FAMILYE1

From equilibrium to a magnetic eigenproblem

A mode perturbs the equilibrium only in directions perpendicular to m₀.

m (r, t) = m_{0} (r) + Re [δ m (r) e^{+ i ω t}], m_{0} \cdot δ m = 0

L q = λ B_{α} q, λ = i ω, δ m = T q

Physical content of L

Linearized exchange, anisotropy and applied/static-field response.
Dynamic demagnetization δH_d[δm], when included.
Projection onto the local tangent plane: two unknown components per active spin.

Meaning of the symbols

T: local tangent basis; q: complex mode coefficients.
L: linearized dynamics operator; B_α: projected mass/gyrotropic operator.
ω: angular frequency [s⁻¹]; f = Re(ω)/(2π) [Hz].

The equilibrium must be stationary enough for “small oscillation around m₀” to be a meaningful approximation.

FULLMAG · EIGENSOLVE · SEPARATE SOLVER FAMILYE2

Boundary conditions define which modes are allowed

Free, pinned and Bloch-periodic boundaries represent different physical experiments.

Condition on δm	Meaning	Effect
Free: ∂ₙδm = 0	Natural exchange boundary without additional surface torque	Boundary spins can precess
Pinned: δm = 0	Dynamic motion suppressed on the selected boundary	Changes standing-wave profiles and eigenfrequencies
Periodic at k = 0	Equal values on corresponding boundary points	The same oscillation repeats in each cell
Bloch/Floquet	δm_dst = exp(−ik·Δr) δm_src	Neighbouring cells have a controlled phase shift; scanning k gives dispersion

q_{d s t} = e^{- i k \cdot Δ r} (T_{d s t}^{T} T_{s r c}) q_{s r c}

k [rad/m] is the wavevector; Δr [m] is the cell translation. The tangent-frame map matters when local bases differ.

A phase factor on two coefficients is sufficient only when the paired tangent bases coincide.

FULLMAG · EIGENSOLVE · SEPARATE SOLVER FAMILYE3

Dynamic demagnetization must obey the same wavevector

The static demag field sets the equilibrium; its linear response belongs inside the eigen-operator.

Δ δ u = \nabla \cdot (M_{s} δ m) in the magnet, Δ δ u = 0 in air, δ H_{d} = - \nabla δ u

Full 3D Bloch cell

Pair both magnetic and airbox faces.
Apply the same complex phase to δm and δu.
Use opposite outward normals in flux matching.
The outer air boundary still determines exterior truncation.

2.5D cross-section is a different model

Replace the propagation derivative by −ik along an invariant axis.
Requires invariant geometry, material, equilibrium and boundary data.
Do not combine shifted derivatives with full-cell phase constraints: that counts k twice.
This cross-section route is a planned path, not qualified support.

A k=0 potential solve is not a substitute for dynamic demag at nonzero k.

FULLMAG · EIGENSOLVE · SEPARATE SOLVER FAMILYE4

What the eigensolver currently contains

Separate source capabilities, admitted workloads and numerical qualification.

Path	Available in the current source	Boundary of the claim
FEM CPU	Eigenmodes, k-path orchestration, dense small-problem oracles and selected PETSc/SLEPc routes	Gamma/free and explicit Floquet payload gates; general selected windows require Full2x2 and ignored damping
FEM CPU dynamic demag-k	Shared-domain Floquet-airbox assembly boundary	Source-visible CPU bridge; no current-main managed qualification established
FEM GPU	K0 shared-domain PETSc/SLEPc path; narrow K0/no-demag cuSolverDN Kittel route	Not a general GPU modal solver; nonzero-k GPU routes remain gated
FDM CPU / GPU	Time-domain trajectories and postprocessed spectra	No native FDM eigen solver in this contract
Spectrum selection	Targeted eigenpairs near a requested frequency	Nearest modes do not certify a complete frequency window

The architecture targets sparse or matrix-free selected modes; dense diagonalization remains a small-system reference.

55 seconds · 34:13–35:08

K0 shared-domain GPU gate requires double, Full2x2, demag, ignored damping, periodic k0, valid Poisson/pair metadata, positive target and count1–32. Current docs label the broader FEM modal work source_visible/unvalidated. PETSc provides matrix/vector and solver machinery; SLEPc supplies scalable eigenproblem algorithms; cuSolverDN is the narrow dense CUDA linear algebra route.

Implementation audit of current main source; no fresh solve or scientific qualification was run for this presentation. Screenshot evidence has its own source date/commit.

FULLMAG · EIGENSOLVE VALIDATIONE4b

Eigensolve versus the Kittel equation

Uniform-mode analytical benchmark

15 points

FEM CPU · k = 0

5.20 × 10⁻¹⁴

Maximum relative error
reported in the figure

The computed eigenfrequencies overlap the analytical curve.

50 seconds · 35:08–35:58

The figure reports an infinite-film periodic benchmark with FEM CPU k=0, 160×80×10 nm, PBC x/y, pure Neumann β=0 and a mean-zero gauge. It reports 15 points and maximum absolute relative error 5.20e-14. These numbers are transcribed from the supplied figure, not independently recalculated. This is a specific uniform-mode analytical benchmark, not evidence of that accuracy for arbitrary meshes, finite k or all interactions. No material parameters absent from the figure are inferred.

Original figure provenance

Details moved from the slide:

Benchmark configuration

160 × 80 × 10 nm cell; periodic in x and y.
k = 0 uniform mode; FEM CPU.
Pure Neumann, β = 0; mean-zero gauge.

Reported agreement

15 computed points follow the analytical curve.

Maximum relative error:
5.20 × 10⁻¹⁴

The lower panel resolves differences hidden by the overlaid curves.

Author-supplied comparison · frequency versus external field.

OPTIONAL · G15 · SUPPORTING DETAILSkip extras →

Docs analytical validation

The validation page distinguishes local contract checks from runtime qualification in the matching MFEM/CUDA environment.

FULLMAG · EIGENSOLVE · SEPARATE SOLVER FAMILYE5

Dispersion: connect each frequency to its wavevector

A k-path is a sequence of distinct eigenproblems, not a single frequency spectrum.

Recorded FEM dispersion UI with analytical reference curves.

Read the comparison

Horizontal position identifies the wavevector sample.
Each branch follows an eigenmode across samples.
Reference curves test a specified analytical model and boundary assumptions.

Ask before claiming agreement

Do the thickness, material, equilibrium, exchange and exterior conditions match? Track modes consistently and check spatial/solver convergence.

This recorded case comes from its source thread; it is not a fresh execution of today’s main checkout.

A dispersion curve is meaningful only together with the physical assumptions and mode identity.

OPTIONAL · G1 · SUPPORTING DETAILSkip extras →

Dispersion: FEM meets analytical references

Fullmag plots a 15-point FEM eigenmode branch against finite-airbox and open-air reference curves; the reported differences are descriptive comparisons, not convergence or validation proof.

Eigenmode dispersion; ky; 15 FEM points; two reference curves; run run-live-93df870d6af7a8e9d70a3742a0de6e7860b5a072c98b7359c3463c18b49113de.

The Fullmag Results workspace rendered a 15-point FEM eigenmode branch with a finite-Dirichlet-airbox n=0 analytical reference and an open-air 1D reference. The run report records maximum relative differences of 0.8271% and 0.5936%, respectively.

Limits: The UI labels scientific trust UNKNOWN. The comparison report says it is descriptive and does not establish mode matching, spectrum completeness, or convergence. The associated browser smoke ended in a 60 s wait timeout after saving this screenshot.

Source: /mnt/c/Users/Mateusz/.codex/visualizations/2026/10/26/01a1770e-c06b-7282-9514-c818c5e8dfc1/eigensolve-ui-recovery-20261028/after-host-restart/dispersion-ui.png

Thread: 01a1770e-c06b-7282-9514-c818c5e8dfc1 · Commit: 7abc39f3ba46a677a3dda12d0493004c87cf7eb6

Full screenshot provenance

OPTIONAL · G5 · SUPPORTING DETAILSkip extras →

Damon–Eshbach dispersion: a diagnostic comparison

Diagnostic Damon–Eshbach dispersion: 15 FEM points versus an open-film analytic curve; the figure itself marks scientific qualification NOT VERIFIED.

FULLMAG · EIGENSOLVE · SEPARATE SOLVER FAMILYE6

Mode shape: a complex spatial oscillation

A frequency alone does not tell us whether a mode is edge-localized, uniform or confined across the waveguide.

Supplied Fullmag mode view: sample selection, component/phase controls and spatial field display.

What to inspect

Select the wavevector and mode index.
Inspect real, imaginary, magnitude or phase representation.
Follow nodes, localization and relative phase across the sample.
Keep display gain distinct from physical amplitude.

Animation

Reconstruct Re[δm exp(iωt)]. The eigenvector normalization is arbitrary until an amplitude convention or driven response is specified.

Colour and phase controls reveal the mode; they do not change the eigenproblem that produced it.

OPTIONAL · G23 · SUPPORTING DETAILSkip extras →

Eigenmode visualization in Fullmag

Select a mode, inspect its spatial representation and adjust phase/display controls.

OPTIONAL · G2 · SUPPORTING DETAILSkip extras →

Inspect a mode in three dimensions

Fullmag displays eigenmode sample 7, mode 0 (9.299 GHz) in a 3D FEM view with real-component controls; this UI capture does not qualify hardware rendering or numerical convergence.

Eigenmode sample 7, mode 0, 9.299 GHz; phase-rotated real display; run run-live-93df870d6af7a8e9d70a3742a0de6e7860b5a072c98b7359c3463c18b49113de.

The Fullmag UI selected eigenmode sample 7, mode 0, and rendered its real-component 3D view with Inspector controls. The accompanying browser report passed and recorded the mode field resource as ready.

Limits: This is UI evidence for one mode on the shown FEM geometry. The report scopes the capture to published FEM artifacts and explicitly says SwiftShader rendering is not hardware-GPU or convergence qualification.

Source: /mnt/c/Users/Mateusz/.codex/visualizations/2026/10/26/01a1770e-c06b-7282-9514-c818c5e8dfc1/eigensolve-ui-recovery-20261028/after-host-restart/gamma/report-real.png

Thread: 01a1770e-c06b-7282-9514-c818c5e8dfc1 · Commit: 7abc39f3ba46a677a3dda12d0493004c87cf7eb6

Full screenshot provenance

OPTIONAL · G4 · SUPPORTING DETAILSkip extras →

Connect a spectral point to its spatial mode

A wider Fullmag 3D view of eigenmode sample 7, mode 0 (9.299 GHz) on the FEM mesh; this is a single-mode UI capture, not convergence evidence.

OPTIONAL · G10 · SUPPORTING DETAILSkip extras →

Complex mode controls · UI fixture

Fullmag’s mode Inspector fixture shows complex representation and phase-playback controls; playback rate is visual only.

OPTIONAL · DEMO · SUPPORTING DETAILSkip extras →

From a dispersion point to a mode

Follow the recorded views; local demonstration servers are not part of this public presentation.

Open recorded dispersion view ↗Open recorded 3D mode ↗Show recorded screenshots →

Inspect the dispersion branch and reference assumptions.
Select a point; show the mode, component and phase controls.
Replay the saved YIG field; explain the recorded run status.

FULLMAG · EIGENSOLVE · SEPARATE SOLVER FAMILYE7

Time-domain FFT and eigensolve provide complementary spectra

The bright lines in an FFT depend on how the sample was excited and observed.

Supplied YIG 2D FFT: frequency versus kₓ, relative power in dB.

Time-domain analysis

Excite the system; record m(x,t).
Apply declared spatial/time windows and Fourier transform.
Finite time and length set spectral resolution; drive overlap sets visible mode strength.

Eigenmode analysis

Find the natural frequencies and spatial eigenvectors directly around equilibrium. A mode can exist even if the chosen antenna couples to it weakly.

A missing bright FFT branch does not prove that the corresponding eigenmode does not exist.

OPTIONAL · G25 · SUPPORTING DETAILSkip extras →

YIG spin-wave dispersion from a 2D FFT

Frequency–wavevector power map from the supplied YIG analysis.

FULLMAG · METHODS & IMPLEMENTATIONF22

Control Room: follow the scientific object through the workflow

One workspace connects model construction, solver selection and published results.

Recorded Fullmag workspace: model tree, central viewport and object inspector.

Three complementary views

Explorer: geometry, materials, stages and result products.
Viewport: mesh, magnetization and selected physical fields.
Inspector: parameters, display controls and result metadata.

Reusable experiment

Python and the UI converge on the same problem description. A saved result should remain inspectable after the solver has stopped.

The UI should explain the model and evidence, not hide which numerical method or field is selected.

OPTIONAL · G6 · SUPPORTING DETAILSkip extras →

Build the model: geometry, regions & materials

Fullmag’s light-theme workspace shows a mixed-P1 model, Airbox, and selected box material/region Inspector before Compute.

OPTIONAL · G7 · SUPPORTING DETAILSkip extras →

The same workspace in dark mode

The same Fullmag mixed-P1 setup in dark theme, showing Airbox and object/material controls before Compute.

OPTIONAL · G12 · SUPPORTING DETAILSkip extras →

Docs control room workspace

The Control Room guide describes the ribbon, Explorer, viewport, Inspector, and status regions. Its embedded workspace illustration visibly contains API-version/resource-load errors, so this is documentation layout evidence only.

OPTIONAL · G16 · SUPPORTING DETAILSkip extras →

Docs ui workspace overview

The full-resolution workspace illustration shows the Explorer, viewport, Inspector, toolbar, and status area; it visibly includes API contract mismatch and resource-load errors.

OPTIONAL · G17 · SUPPORTING DETAILSkip extras →

Docs ui ribbon geometry

The published Geometry ribbon strip shows the application header and geometry-tab placement; it is a narrow navigation illustration, not evidence that geometry commands executed.

OPTIONAL · G19 · SUPPORTING DETAILSkip extras →

Docs ui inspector draft

The published Inspector panel image illustrates the property-editor region and shows resource-load failure notices.

OPTIONAL · G13 · SUPPORTING DETAILSkip extras →

Docs control room viewport

The guide describes the 3D Viewport and shows its empty grid, orientation cube, and axis gizmo; the illustration does not show an active magnetization field.

OPTIONAL · G18 · SUPPORTING DETAILSkip extras →

Docs ui viewport field

The guide’s viewport image shows the 3D canvas, axis orientation, and API-version mismatch/resource-load errors; no magnetization field is rendered.

OPTIONAL · G20 · SUPPORTING DETAILSkip extras →

Docs ui status footer

The public guide’s status/footer strip shows the application’s lower status area; it is only a layout illustration.

FULLMAG · METHODS & IMPLEMENTATIONF23

A useful status map: what works, what is being extended

Implementation breadth is real; qualification must remain attached to a named numerical workload.

Area	Current source / demonstrated material	Development focus
FDM dynamics	Rust CPU reference; CUDA single-grid FP32/FP64; supplied MuMax3 comparison	Controlled error metrics and broader matched cases
FDM multilayer	Ordered-pair convolution, native geometries and CPU transfers; CUDA components	Descriptor parity, unequal-thickness end-to-end cases, device qualification
FEM demag	Airbox Poisson and hierarchical FK CPU/GPU implementations	Per-realization accuracy, exterior convergence and managed GPU evidence
Interaction coverage	Conservative fields, rotated DMI, bounded torques/current routes	Close workload-specific gaps rather than claim every combination
Eigensolve	Separate FEM linearized solver family with boundary/k-path machinery	Nonzero-k dynamic demag, native sparse paths and mode qualification
Workflow	Python / UI authoring, saved data, dispersion and mode inspection	Make method, units, boundary conditions and provenance visible

Fullmag’s value is the connection between new physical models, numerical choices and inspectable results.

FULLMAG · METHODS & IMPLEMENTATIONF24

What we can do together with Fullmag

Choose a scientific question, then choose the numerical experiment that can answer it.

Geometry and micromagnetics

Use FDM for efficient regular films; FEM for geometry-driven questions.
Test exchange, anisotropy, demag and chiral terms—including rotated DMI.
Compare a controlled case with MuMax3 or a matching FEM formulation.

Excitation and magnonics

Compute/inspect a spatial antenna field.
Study nonlinear trajectories in time.
Use the separate eigensolver to connect dispersion to mode shape.
Compare mode predictions with the spectrum excited by a given drive.

Next scientific deliverable: one reproducible geometry, named boundary conditions, selected solver/device, archived inputs/results and quantitative comparison.

The purpose is a research platform we can extend and understand—not a blanket replacement for established tools.

Presentation available:
presentations.mzelent.pl
/28102026/ai.html

AI in Research& FullMag

Practical AI workflows.A platform for micromagnetic simulation.

Claude · model → Effort

ChatGPT / Work · model + reasoning

Where to set them

A useful researcher preset

What if these were GPT-6.1 Sol API tokens?

Worth remembering

Keep in versioned files

Example: manuscript claim audit

What goes in the skill

1 · Search and inspect

2 · Choose one skill

Prompt to copy

What the skill adds

1 · Explore in Chat

2 · Export the evidence

3 · Continue with files

Before leaving Chat

First prompt in VS Code

A practical loop

Source ZIP · broadest access

Git · when enabled for the project

VS Code extension

CLI · native install

Terminal · before the session

Claude Code · after it opens

Starter example

Write rules that can be checked

Create this file

Discover and invoke

Inspect what changed

Compile and inspect the PDF

ZIP → new Overleaf project

Git → confirmed remote + branch

Attach an evidence pack

Request a bounded change

Bad prompt

Attach

Good prompt

Illustrative API calculation

What increases actual usage

01 · Scope the context

02 · Bound the output

03 · Keep a handoff

04 · Route by difficulty

05 · Reuse evidence

06 · Measure the loop

Explore new physics

Choose the right method

Keep the experiment connected

An idea becomes a project

AI-assisted development

A shared research framework

What the terms do

Symbols and units

FDM: neighbours and cell faces

FEM: basis functions and integrals

Meaning and authoring

Implemented scope

Operator and algorithm

What runs where

What exists

What is still being completed

Numerical sequence

Outer boundary is an approximation

Implementation choices matter

Comparisons need the same history

Read the branches

Interpretation

Time-domain LLG

Eigenmode problem

AI in Research
& FullMag

Practical AI workflows.
A platform for micromagnetic simulation.