AI for physics: a landscape review
AI is not replacing physicists — but it is compressing the timescale of structured research. The evidence now extends beyond benchmark results to peer-reviewed publications: a Harvard QCD expert used Claude Opus 4.5 as a research collaborator to take a paper from conception to arXiv in approximately two weeks, a task that would typically take one to two years. The AI contribution was significant but bounded — it handled well-defined calculation steps under constant expert supervision. This is the defining pattern across all current AI-for-physics tooling: powerful acceleration of structured, well-posed problems; limited utility for genuinely novel conceptual leaps.
Three emerging paradigms
1. Human-in-the-loop LLM collaboration — the Schwartz/Anthropic model
What it is: A domain expert coaches a general-purpose LLM through a complex, multi-step physics calculation, reviewing each step and correcting errors before proceeding.
What it has achieved: Harvard professor Matthew Schwartz used Claude Opus 4.5 + Claude Code (VS Code extension) to produce a full quantum chromodynamics paper on the C-parameter Sudakov shoulder (arXiv 2601.02484). The workflow generated 51,248 messages and approximately 36 million tokens. Total human oversight was 50–60 hours. GPT 5.2 and Gemini 3.0 were used for cross-verification of key results. The paper was submitted to arXiv; the workflow is documented in Anthropic’s “Vibe Physics” research blog post.
What it cannot do: The AI was characterised as performing at the level of a “second-year graduate student” — competent at executing well-defined calculations but prone to fabricating results and struggling with honest verification. It cannot identify which calculation to do, only do a calculation that an expert has already scoped. Every AI output required domain-expert review before being trusted.
Accessibility: Available now via Claude API / Claude Code subscription and comparable LLM services. No specialist tooling required beyond the model itself. Suitable for researchers with significant domain expertise who can provide constant oversight.
2. Structured multi-agent orchestration — Get Physics Done
What it is: An open-source agent framework that decomposes a physics research problem into four structured phases (Formulate → Plan → Execute → Verify) and coordinates specialist sub-agents through each phase, with an automated physics verification layer.
What it has achieved: get-physics-done automates checks for dimensional consistency, limiting cases, symmetry constraints, conservation laws, and numerical stability — the kind of routine verification an expert should do but often skips under time pressure. It produces arXiv-ready LaTeX manuscripts, Python verification scripts, and figures as end-to-end output. It also integrates with arXiv via arxiv-mcp-server for literature retrieval.
What it cannot do: GPD does not generate novel physical intuition. Its verification layer is bounded by known constraints; genuinely unprecedented problems will not be well served. The tool is too new (released March 2026) for independent benchmarking of its failure modes. Community commentators note that it can follow physics rules convincingly but may struggle with truly novel conceptual leaps — consistent with the broader pattern from the Schwartz case study.
Accessibility: Apache 2.0, free to install (npx -y get-physics-done or pip install get-physics-done), but requires a paid LLM runtime backend (Claude Code, Gemini CLI, Codex, or OpenCode). API costs apply. Suitable for long-horizon structured computational and theoretical physics projects; lighter tasks may not justify setup overhead.
3. Autonomous laboratory — emerging but not yet ready
What it is: AI systems that propose and run wet-lab or simulation experiments without continuous human intervention.
What has been achieved: Tools such as google-co-scientist have demonstrated autonomous hypothesis generation and literature synthesis for life sciences. Purpose-built autonomous lab systems (Coscientist and similar) have demonstrated self-directed chemical synthesis for simple reactions. None of these have been demonstrated for high-energy or theoretical physics at research frontier level.
What it cannot do: Genuine autonomous frontier research in physics — the design of novel experiments, the identification of non-obvious theoretical frameworks, the judgement of what constitutes an interesting result — remains beyond demonstrated capability. Current autonomous systems excel at well-defined search problems within known solution spaces.
Accessibility: No generally accessible autonomous physics research tools exist at time of writing (2026-Q1). This is a space to monitor over 12–24 months.
What the Schwartz case study tells us
The vibe-physics case study (arXiv 2601.02484; Anthropic blog post, 2026) is the most detailed public account of LLM-assisted physics research to date. Key takeaways for institutional planning:
| Dimension | Finding |
|---|---|
| Speed | ~2 weeks vs typical 1–2 years for a comparable QCD calculation |
| Human cost | 50–60 hours of expert oversight — not “autonomous” |
| AI capability level | ”Second-year graduate student” |
| Primary failure mode | Fabrication of results; poor self-verification |
| Mitigation | Cross-verification with multiple LLMs (GPT 5.2, Gemini 3.0); expert review of every step |
| Tool stack | Claude Opus 4.5 + Claude Code (VS Code) |
| Output | Peer-reviewed arXiv paper (2601.02484) |
The core finding: AI acceleration is real and substantial, but it is not a substitution for domain expertise. The workflow requires a physicist who already knows the answer is reachable, can scope each calculation step, and can verify each output. Junior researchers or non-specialists cannot replicate this workflow without the expert oversight layer.
Relevance to Sheffield
Who benefits now
| Audience | Tool / approach | Why |
|---|---|---|
| Theoretical / computational physicists | LLM collaboration (Claude / GPT) | Speed gains on well-scoped derivations; manuscript drafting |
| Theoretical / computational physicists | get-physics-done | Structured workflow with automated verification layer |
| Physics PhD students | LLM collaboration | Accelerates routine calculation steps; frees supervisor time for harder problems |
| Computational Engineering | get-physics-done | Physics-grade verification for simulation and modelling projects |
Who should monitor (12–24 months)
- Experimental physics groups — as AI-assisted data analysis and simulation tooling matures
- Materials science and condensed matter — close overlap with the QCD calculation workflow; similar calculation-heavy structure
- Physics educators — LLM-as-graduate-student has direct implications for how problem sets and research supervision are designed
Barriers to adoption
- Expert oversight requirement — the Schwartz workflow required 50–60 hours of senior expert time. This is a significant real cost, even if the speedup is dramatic. Researchers need to account for this in project planning.
- Fabrication risk — both the Schwartz case study and community GPD commentary confirm that LLMs fabricate physics results with apparent confidence. Any AI-assisted physics workflow requires a verification step that cannot itself be delegated to the AI.
- Data governance — both the LLM collaboration and GPD approaches involve sending research content to third-party US-hosted cloud APIs. Unpublished results, sensitive simulation data, and commercially sensitive material should not be used without a University DPA review of the specific backend.
- Cost at scale — the Schwartz paper consumed ~36M tokens. At current Claude Opus pricing, the API cost for a comparable project would be substantial. Researchers should model token costs before committing to this workflow for a major paper.
Recommended actions
Immediate (now)
-
Share the Schwartz/Anthropic case study with Physics and Applied Mathematics — arXiv 2601.02484 and the Anthropic vibe-physics blog post together provide a compelling and credible account of what is possible today. Most researchers will find the “second-year graduate student” framing useful for calibrating expectations.
-
Point experimental and computational physics researchers to get-physics-done — it is free, available now, and provides the structured workflow and automated verification layer that makes LLM-assisted physics research more tractable for researchers without Claude/GPT expertise.
Near-term (Q2–Q3 2026)
-
Commission a token-cost analysis — model the API cost of a Schwartz-scale project (36M tokens) against current Claude and Gemini pricing, and assess whether a departmental API key arrangement would be cost-effective relative to individual researcher subscriptions.
-
Commission a GDPR/DPA review — before formally recommending LLM-assisted physics workflows for research involving unpublished data, Research IT / IT Services should review the data processing terms for Claude API and Gemini API. The Schwartz paper used proprietary QCD calculations; Sheffield researchers will face the same question.
-
Run a pilot — identify one or two PhD students or postdocs in theoretical or computational physics who are willing to test the Schwartz workflow on a bounded calculation. A structured pilot with a debrief report would provide Sheffield-specific evidence beyond the Harvard case study.
Strategic (2027+)
- Revisit autonomous physics tools as the space matures — current autonomous lab and hypothesis-generation tools are not yet reliable for theoretical physics, but the pace of progress in adjacent domains (chemistry, biology) suggests this will change.
Cross-references
Tool notes in this catalogue covering the AI-for-physics space:
- get-physics-done — structured multi-agent physics research; open-source;
assess - google-co-scientist — general scientific hypothesis generation; stub only
- paperqa2 — literature-grounded scientific QA; relevant for physics literature review
- ag2-autogen — general multi-agent orchestration; relevant infrastructure layer
See also: ai-for-mathematics — overlapping tooling (GPD appears in both); complementary framing.
Sources
- Anthropic “Vibe Physics” research blog post: https://www.anthropic.com/research/vibe-physics (2026)
- arXiv 2601.02484 — Schwartz et al.: C-parameter Sudakov shoulder paper produced via vibe-physics workflow
- Get Physics Done GitHub repository: https://github.com/psi-oss/get-physics-done
- Physical Superintelligence announcement (Substack): https://theinnermostloop.substack.com/p/physical-superintelligence
- Anthropic model documentation: Claude Opus 4.5 capabilities and context window
Drafted: 2026-03-25 · Branch: agent/2026-03-25-vibe-physics-incorporation · Requires human review before promotion to main