# AI practice and Fullmag presentation plan

**Working title:** AI in Research and the Fullmag Simulation Platform  
**Subtitle:** From useful assistance to traceable scientific computation  
**Language:** English  
**Audience:** departmental researchers; introduce physical and numerical concepts before implementation detail  
**Duration:** 25 minutes, excluding questions  
**Structure:** 8 AI slides in 10 minutes; 12 Fullmag slides in 15 minutes; technical backup slides outside the speaking time  
**Prepared:** 28 October 2026

This is a content and design specification for the first preparation stage. The main slides carry the argument; the accompanying research briefing and technical dossier carry implementation detail. A complete code listing would not be readable in a 15-minute scientific introduction. Instead, the main talk explains the full path from physical problem to numerical result, and the backup material maps it to source owners and qualification limits.

## Narrative

The audience should leave with three conclusions: AI tasks need an evidence contract; Fullmag aims to make numerical choices explicit within one scientific workflow; and the value of a simulation platform depends on demonstrated physics and reproducibility as well as interface quality.

The transition is deliberate: **“The same question applies to AI and simulation software: what evidence makes this result trustworthy?”**

## Visual direction

Use a restrained editorial style with generous whitespace, precise diagrams and large equations. The deck should feel like a scientific seminar, not a product launch.

| Element | Specification |
|---|---|
| Canvas | 16:9 widescreen; approximately 7% outer margins |
| Background | Warm white #F6F7F5; dark graphite #17212B for the opening and Fullmag section start |
| Text | Graphite #17212B; muted slate #596675 for secondary information |
| Section accents | Teal #087F8C for AI; blue #365CF5 for Fullmag |
| Warning/uncertainty | Amber #A46110, always paired with a word or symbol |
| Typography | Aptos or another installed neutral sans serif; Cambria Math for editable equations; monospaced source labels only when necessary |
| Scale | Titles 34–40 pt; main text 24–28 pt; equation 28–36 pt; figure labels at least 18 pt; source footer at least 12–14 pt |
| Density | Usually one message, one visual and at most three short text items; about 25–45 on-slide words excluding labels |
| Motion | Simple appearance builds for a workflow or equation; no spinning objects, animated backgrounds or live typing |
| Citations | A short source label on the slide; complete link and source/version information in notes |
| Status | Use text labels “implemented in source”, “qualified in this case”, “partial” and “not verified”; no unexplained green checks |
| Comparisons | Small MuMax3 and COMSOL strip on each Fullmag slide; use comparable scope, not a winner/loser score |

Use consistent visual grammar. Teal arrows represent information flow; blue elements represent numerical operations. CPU/GPU and FDM/FEM must be named, not encoded by color alone. Physical vectors and mesh edges should remain visible on a projector. Do not use generated pseudo-scientific field images as simulation evidence.

## Timed running order

| # | ID | Slide title | Time | Cumulative |
|---|---|---|---|---|
| 1 | A1 | AI in Research and the Fullmag Simulation Platform | 0:30 | 0:30 |
| 2 | A2 | Three ecosystems, one verification standard | 1:25 | 1:55 |
| 3 | A3 | Choose the working environment for the task | 1:05 | 3:00 |
| 4 | A4 | Give the model a contract and an evidence pack | 1:15 | 4:15 |
| 5 | A5 | Skills turn a good procedure into a reusable workflow | 1:10 | 5:25 |
| 6 | A6 | Literature research starts with an evidence table | 1:35 | 7:00 |
| 7 | A7 | Improve the manuscript without inventing the science | 1:30 | 8:30 |
| 8 | A8 | Spend stronger reasoning on a well-defined question | 1:30 | 10:00 |
| 9 | F1 | Why we are building Fullmag | 0:55 | 10:55 |
| 10 | F2 | What micromagnetics actually computes | 1:05 | 12:00 |
| 11 | F3 | Energy defines the field; LLG defines the motion | 1:25 | 13:25 |
| 12 | F4 | The same magnet can become a grid or a mesh | 1:05 | 14:30 |
| 13 | F5 | FDM maps regular structure onto efficient kernels | 1:20 | 15:50 |
| 14 | F6 | FEM represents geometry through a weak formulation | 1:25 | 17:15 |
| 15 | F7 | An interaction has four implementation lanes | 1:10 | 18:25 |
| 16 | F8 | One model flows through a layered architecture | 1:20 | 19:45 |
| 17 | F9 | Each technology has a specific numerical or product role | 1:15 | 21:00 |
| 18 | F10 | What is implemented, and what is qualified | 1:25 | 22:25 |
| 19 | F11 | Compare methods at matched accuracy and scope | 1:25 | 23:50 |
| 20 | F12 | The value for our group, and the next evidence we need | 1:10 | 25:00 |

## AI slide specifications

### A1 — AI in Research and the Fullmag Simulation Platform

**On slide:** “AI practice · 10 min” and “Fullmag · 15 min”; one sentence: “Useful outputs need traceable evidence.”

**Visual:** a quiet title slide with two aligned section markers; a small original vector sketch moving from document to numerical field.

**Speaker cue:** “I will start with practical ways to use AI in our research. Then I will show why we are building Fullmag, how it represents the physics, and which parts are implemented and verified.”

**Purpose:** set an honest scope before introducing tools.

### A2 — Three ecosystems, one verification standard

**On slide:** three equal columns: “Anthropic / Claude”, “OpenAI / ChatGPT”, “Google / Gemini”. Under each: one research workflow, one useful capability, one limitation.

**Visual:** a three-column comparison, not vendor logos dominating the page. Use the task matches in the AI briefing.

**Speaker cue:** Explain that all three can assist research and all need verification. Choose according to available sources, execution tools, integration and the quality of outputs on the group's own tasks. Separate a provider from its models and subscription. Avoid universal rankings and benchmark scores unrelated to the audience's work.

**Sources:** AI briefing, “Three ecosystems and their tradeoffs”; official Claude Research, ChatGPT Work/Codex and Gemini Deep Research guides.

### A3 — Choose the working environment for the task

**On slide:** “Chat → explanation”, “Work → finished artifact”, “Code → repository change”. Footer: “Access and tools matter as much as the model.”

**Visual:** three task cards: understand a paragraph; deliver a literature matrix; repair an analysis script.

**Speaker cue:** These are workflow orientations, not rigid capability walls. Work can execute code and a coding agent can edit a manuscript. Mention Claude Code and agentic Cowork workflows; note that Claude's current chat/Cowork interface is being unified. The demonstration should follow the actual account interface rather than assume everyone sees the same buttons.

**Sources:** OpenAI Work overview and current experience guide; Claude experience update.

### A4 — Give the model a contract and an evidence pack

**On slide:** “Task · Context · Sources · Constraints · Output · Checks”. A short example: “Edit the Introduction. Preserve claims and citation keys. Return a diff and uncertainties.”

**Visual:** a prompt card on the left; four named attachments on the right. One arrow joins them.

**Speaker cue:** Compare this request with “improve my paper”. Explain why a bibliography, numerical table and figure caption serve different roles. Ask the model to confirm what it could read. An unreadable scanned equation or a missing figure is a context problem, not a reason to buy more reasoning.

**Sources:** AI briefing, prompt and attachment examples; current upload guides.

### A5 — Skills turn a good procedure into a reusable workflow

**On slide:** “Instructions + references + optional scripts”. Example: “Manuscript claim audit”. Footer: “A procedure, not a newly trained model.”

**Visual:** three simple folder-shaped cards feeding a claim/evidence/result table. Show only four rows of a sample output.

**Speaker cue:** Define the skill with the laboratory SOP analogy. Distinguish a skill from a project context and a connector. Explain that scripts and tool access still have permissions. Gemini Gems offer reusable instructions and knowledge, but should not be described as automatically identical to portable Agent Skills.

**Sources:** OpenAI and Claude skill documentation; Gemini Gems documentation.

### A6 — Literature research starts with an evidence table

**On slide:** “Question → Search → Screen → Extract → Verify”. Small evidence-table example with DOI, method, finding and source location.

**Visual:** an original five-stage line, with verification made visually explicit. Label example records “template” unless populated from real checked papers.

**Speaker cue:** Ask for full-text access status, distinguish preprints from final publications, and verify DOI metadata and the cited passage. Research tools accelerate discovery; they do not establish exhaustive coverage. A systematic review requires a documented search and screening protocol. The final bibliography must come from verified records.

**Sources:** AI briefing literature workflow and official research guides. No fabricated papers in mockups.

### A7 — Improve the manuscript without inventing the science

**On slide:** “Separate structure, science, language and references.” “Preserve a versioned original.” “Review every substantive change.”

**Visual:** a small before/after paragraph with one legitimate clarity edit and one highlighted overclaim to reject. Use an original invented teaching sentence clearly labeled “illustrative”.

**Speaker cue:** Show how “is consistent with” can be incorrectly turned into “demonstrates”. A polished revision can change scientific meaning. Use AI to audit symbol consistency and reviewer responses, but keep authors accountable. Follow the current journal policy and use approved environments for unpublished or confidential material.

**Sources:** AI briefing manuscript prompts; Nature AI policy, with target-journal policy rechecked before the talk.

### A8 — Spend stronger reasoning on a well-defined question

**On slide:** “Organize first → reason deeply → verify independently”. Footer: “Pro subscription ≠ selected model ≠ research mode.”

**Visual:** a three-step ladder with a derivation as the difficult middle task and unit/limit checks as the last step.

**Speaker cue:** Use a stronger model for a bounded derivation or a difficult review, after supplying sources and conventions. A premium model cannot recover an absent measurement. End with: “The same question applies to AI and simulation software: what evidence makes this result trustworthy?”

**Sources:** AI briefing Pro section and current OpenAI model guide. Avoid pricing and quota details in the live talk.

## Fullmag slide specifications

### F1 — Why we are building Fullmag

**On slide:** “One scientific model.” “Explicit numerical choices.” “Traceable results.”

**Visual:** a physical sample in the center, connected to a regular grid, a conforming mesh, a script and an analysis view.

**Speaker cue:** Explain the group's need to move between fast structured-grid micromagnetics, complex geometry and richer experiments without losing the meaning of the model or its provenance. Fullmag's motivation is control of this workflow and extensibility for research. Present those as design goals, not completed superiority claims.

**Comparison strip:** “MuMax3: focused GPU micromagnetics · COMSOL: broad FEM/multiphysics · Fullmag: a common explicit workflow across numerical lanes.”

**Evidence:** technical dossier motivation and architecture; external comparison register.

### F2 — What micromagnetics actually computes

**On slide:** reduced magnetization m = M/Ms, unit length; “Exchange · Anisotropy · Demagnetization · Applied field”.

**Visual:** a few magnetic arrows with four small physical effects. Use the same sample as F1.

**Speaker cue:** Explain a continuum magnetization field rather than individual atomic spins. Exchange favors smoothness, anisotropy favors directions, magnetostatics couples distant regions, and an applied field drives the system. State the fixed-magnitude continuum assumption before discussing more elaborate interactions. The cell or element size must resolve the relevant physical length scales.

**Comparison strip:** “All three can represent micromagnetic models; the physical equations must be matched before comparing discretizations.”

**Evidence:** dossier equations and assumptions; MuMax3 design paper; COMSOL micromagnetics article.

### F3 — Energy defines the field; LLG defines the motion

**On slide:** two builds: the energy-to-field relation, then the explicit LLG equation. Add short labels “precession”, “damping”, “direct torque”.

**Visual:** large editable equations from the dossier, labeled eq-talk-effective-field and eq-talk-llg. Keep the complete energy expansion and symbol table in backup B4.

**Speaker cue:** The effective field is derived from the energy; dynamics combines precession with damping. Direct spin torques enter the time derivative separately. Fullmag uses H in A/m and gamma_mu0 already contains mu0. MuMax3 exposes many fields as B in tesla, so equal-looking numbers are not sufficient for a fair comparison.

**Comparison strip:** “Match H/B, gamma, signs, damping and boundary conditions.”

**Evidence:** Fullmag units and LLG source owners in the dossier; MuMax3 API field units.

### F4 — The same magnet can become a grid or a mesh

**On slide:** “FDM: regular cells and local stencils”. “FEM: elements, basis functions and weak forms”.

**Visual:** original side-by-side cutaways of the same curved magnetic body. Grid and mesh must show the same dimensions and boundary.

**Speaker cue:** FDM is attractive for regular geometry and efficient convolution; curved boundaries require discretization care. FEM can represent boundaries and local mesh variation more naturally, but introduces mesh construction, integration and solver complexity. Neither method is automatically more accurate. Convergence determines accuracy.

**Comparison strip:** “MuMax3 → structured FDM · COMSOL → FEM framework · Fullmag → explicit choice of both.”

**Evidence:** dossier numerical-method sections; primary MuMax3 and COMSOL references.

### F5 — FDM maps regular structure onto efficient kernels

**On slide:** “Local fields → FFT demagnetization → time step → accepted state”. One discrete exchange stencil or convolution equation.

**Visual:** a four-step data-flow with a bracket showing the resident GPU working set. Label CPU and GPU realizations separately.

**Speaker cue:** Explain neighbor-based local terms and the nonlocal magnetostatic convolution. Identify the actual Rust/CPU and CUDA/cuFFT paths from the dossier. FFT workspace, precision, boundaries and output transfers matter. Having CUDA source is not the same as a qualified run for every interaction and geometry.

**Comparison strip:** “MuMax3 is the relevant FDM reference; COMSOL's GPU sparse solver is a different workload.”

**Evidence:** dossier FDM source map and qualification limits.

### F6 — FEM represents geometry through a weak formulation

**On slide:** magnetization basis expansion plus one exchange weak-form relation. A separate label: “Magnetostatics needs an exterior-field treatment.”

**Visual:** element/basis-function sketch on the left; magnetic region and surrounding air or boundary representation on the right.

**Speaker cue:** Replace pointwise derivatives by integrals against basis/test functions. Explain mass weighting and why magnetostatics cannot be solved by ignoring the space outside the magnet. Identify MFEM and the selected solver stack; distinguish the airbox Poisson and boundary-coupled formulations and their implementation limits. Keep alternative formulations in B6.

**Comparison strip:** “COMSOL is the relevant FEM/multiphysics comparison; MuMax3 provides a matched physical reference where geometries overlap.”

**Evidence:** dossier FEM weak forms, native owners and demagnetization routes.

### F7 — An interaction has four implementation lanes

**On slide:** four columns — FDM CPU, FDM GPU, FEM CPU, FEM GPU — and at most six grouped rows. Use a subset of the full dossier matrix.

**Visual:** readable support table with distinct labels for source implementation, partial support and rejected combinations. Do not compress all interactions onto one slide.

**Speaker cue:** A Python constructor does not prove that all four backends execute it. Follow an interaction through authoring, IR, planner and native implementation. Use exchange as the straightforward example and transport/thermal/eigensolve restrictions as examples where the path matters. State which limits are currently explicit in code.

**Comparison strip:** “Compare the same interaction, parameterization and execution mode in each solver.”

**Evidence:** full interaction matrix in dossier; backup B5.

### F8 — One model flows through a layered architecture

**On slide:** “Python study → ProblemIR → plan → native execution → artifacts → Control Room”.

**Visual:** a horizontal architecture diagram with four solver lanes below the plan. Add a thin feedback path from accepted artifacts to analysis; do not imply the browser runs the solver.

**Speaker cue:** Explain each boundary and why it exists. Authoring expresses intent; planning chooses supported execution; native engines compute; the artifact layer records accepted results; the UI inspects them. Requested GPU and resolved execution must remain visible. The architecture is valuable only if these contracts hold through the entire path.

**Comparison strip:** “MuMax3: compact simulation scripting · COMSOL: integrated model/study framework · Fullmag: typed cross-layer model and execution contract.”

**Evidence:** architecture source map in dossier.

### F9 — Each technology has a specific numerical or product role

**On slide:** four stacks: “Authoring”, “Orchestration”, “Numerics”, “Interface”. Show technology names, not a wall of version numbers.

**Visual:** Python; Rust; C++/CUDA with MFEM and numerical libraries; TypeScript/React/Next.js with Three.js. Place Gmsh near mesh preparation and FFT libraries near FDM.

**Speaker cue:** State what each dependency actually does. Distinguish used runtime libraries from optional build dependencies and future routes. Explain why a JavaScript frontend and a Python API do not mean the core numerical loops execute in those languages. Put exact declared versions and optionality in backup B8.

**Comparison strip:** “Compare solver architecture and workflow; language choice alone does not establish performance.”

**Evidence:** versioned manifest inventory in dossier.

### F10 — What is implemented, and what is qualified

**On slide:** four lane cards with three lines each: “source”, “available evidence”, “remaining qualification”. A small separate row for frequency/eigenmode work.

**Visual:** no percent-complete bars. Use the current source snapshot and dated evidence labels.

**Speaker cue:** Present the strongest justified statement for each lane. Explain that source, a successful build, a native run, browser visualization and scientific parity answer different questions. Historical receipts support their recorded inputs and versions. Do not advertise complete FEM GPU/eigensolve qualification from planner or source presence alone.

**Comparison strip:** “Feature availability and demonstrated agreement are separate columns for every solver.”

**Evidence:** dossier maturity table; exact source snapshot and evidence register.

### F11 — Compare methods at matched accuracy and scope

**On slide:** “Same physics · Same observable · Converged resolution · Recorded hardware”. Show three columns: Fullmag, MuMax3, COMSOL.

**Visual:** comparison table plus an empty benchmark layout labeled “validation design” until real comparable data are available. Do not draw plausible-looking curves or speedup bars.

**Speaker cue:** MuMax3 is a focused GPU micromagnetics reference; COMSOL provides a broad FEM/multiphysics environment, with a customized micromagnetics module described by its authors. Current COMSOL also supports GPU acceleration, including cuDSS. A fair benchmark fixes geometry, physics, initial state, conventions, precision and stopping criteria, then demonstrates spatial and temporal convergence before timing.

**Comparison strip:** “No universal winner; choose a numerical route for the research question.”

**Evidence:** external-source register; dossier proposed validation protocol and NIST SP4 reference.

### F12 — The value for our group, and the next evidence we need

**On slide:** “Reusable studies”, “Auditable results”, “New research methods”. Below: “Next: matched validation cases, clearer lane qualification, reproducible examples.”

**Visual:** three simple research outcomes connected to one evidence checklist. Reuse the title-slide motif.

**Speaker cue:** Fullmag offers an extensible common workflow and direct control over scientific implementation. Its costs are architectural complexity, a wider validation matrix and ongoing qualification. End by proposing that the group nominate one representative thin-film case and one geometry/multiphysics case to drive comparisons. These are proposed priorities, not an approved roadmap or promised delivery dates.

**Comparison strip:** “Build where research needs justify ownership; retain MuMax3 and COMSOL as valuable references.”

**Evidence:** dossier tradeoffs, gaps and proposed next milestones.

## Technical backup slides

These slides are outside the 25-minute route. They provide the detailed implementation discussion requested for questions and follow-up.

| ID | Title | Required content | Primary material |
|---|---|---|---|
| B1 | A reusable prompt and attachment pack | Full six-part brief; file manifest; output acceptance | AI briefing |
| B2 | Literature extraction and manuscript audit | Evidence matrix; claim audit; reviewer response template | AI briefing |
| B3 | Current product terminology | Provider/model/mode/subscription; dated availability; Work vs Code | Official sources |
| B4 | Micromagnetic equations and SI conventions | Energy terms, LLG, symbols, H/B and gamma, direct torques | Dossier equations |
| B5 | Complete interaction support matrix | Each interaction across four lanes, restrictions, source owners | Dossier matrix |
| B6 | Demagnetization in FDM and FEM | Tensor convolution, FFT padding, Poisson weak form, boundary treatment | Dossier demag section |
| B7 | Time integration, relaxation and eigenmodes | Physical time vs minimization; accepted steps; linearized eigenproblem; qualification | Dossier workflows |
| B8 | Libraries and declared versions | Technology, role, version source, optionality, runtime verification boundary | Dossier inventory |
| B9 | Python to IR to native implementation | Real scenario path, builder symbols, serialization and plan consumers | Dossier source index |
| B10 | Fullmag versus MuMax3 and COMSOL | Extended fair comparison; custom COMSOL module distinction; GPU scope | Evidence register |
| B11 | Qualification and benchmark protocol | SP4 and analytical cases; convergence; hardware/precision; receipts | Dossier validation |
| B12 | Limitations and research roadmap | Supported subsets, missing proof, future work and decision criteria | Dossier maturity/gaps |

## Asset and rehearsal plan

Create original vector diagrams for the prompt pack, workflow, energy competition, FDM/FEM geometry and layered architecture. Render equations as editable math or verified vector graphics. If a Control Room screenshot is used, capture the selected runtime without initiating a simulation, identify the actual case, and distinguish software rendering from hardware-GPU computation. A screenshot is optional; diagrams alone are sufficient for this first deck version.

Use simulation plots only when their data, units, source commit, parameters and solver receipts are available. A missing comparison plot is replaced by a validation protocol, not synthetic evidence. Complete scientific references belong in notes and the backup reference slide.

Rehearse the eight-slide AI section to exactly 10 minutes. At 17:15, the talk should have completed the FEM introduction; at 21:00, the technology explanation should be finished. If discussion interrupts the talk, compress technology detail into B8 and retain F10–F12. If questions must fit within the 25-minute slot, reserve 3 minutes and shorten the spoken route to 22 minutes before slide production.

## Content acceptance before deck production

- Exactly two timed sections totaling 600 and 900 seconds.
- All requested AI topics have a visible slide and practical material in the briefing.
- Every Fullmag slide has a meaningful MuMax3/COMSOL comparison or comparison condition.
- Main equations use Fullmag's verified units and conventions; all symbols are defined in the dossier.
- Implementation, runtime evidence, scientific validation and roadmap remain visibly separate.
- The current source snapshot and third-party documentation dates appear in notes.
- No unverified speedup, overall parity claim, fabricated plot or unsupported claim of complete backend readiness.
- Exact versions and complete matrices are available in backup material without shrinking main-slide text.