PhysicsSolutionAgent (PSA) System

• PSA is an autonomous, agent-driven system for solving physics problems using integrated LLMs, high-fidelity simulation codes, and visual feedback.
• It employs a modular multi-agent architecture where specialized agents perform tasks like inverse design, simulation, and evaluation under central planning.
• Its hybrid optimization approach combines numerical methods and visual meta-reasoning to achieve rapid convergence and scalable scientific discovery.

PhysicsSolutionAgent (PSA) denotes an autonomous, agent-driven computational system engineered for end-to-end solution, explanation, and analysis of numerical and theoretical physics problems. PSA systems are built around tightly integrated multi-agent architectures, leveraging LLMs, high-fidelity simulation codes, deep learning emulators, and multimodal feedback mechanisms to augment human-level physics reasoning and automate tasks such as numerical simulation, visual explanation, and scientific discovery (Thole et al., 19 Jan 2026, Shachar et al., 2 Oct 2025).

1. Multi-Agent Architectures and Modular Design

PSA frameworks typically adopt a modular, micro-service agent paradigm, decomposing workflow across specialized agents governed by a central planner. The “Multi-Agent Design Assistant” (MADA) exemplifies this approach with a PlanningAgent that orchestrates four specialist agents: InverseDesignAgent (IDA), SimulationAgent, JobManagementAgent (JMA), and Professor (EmulatorAgent). Each agent is based on a fine-tuned LLM and is limited to a restricted set of tools to control behavior and avoid code generation errors (Shachar et al., 2 Oct 2025).

Communication between agents in PSA is structured via strict JSON-RPC–style message envelopes:

{ "agent": "IDA"|"Sim"|"JMA"|"Professor", "tool_call"?: { "name": "...", "args": {...} }, "message": "..." }

Turn-taking and agent invocation are mediated by the PlanningAgent, based on user prompts parsed for target agent, parameter arguments, and desired output formats.

This modular structure establishes clarity, explainability, and robustness, and is recognized as essential for scaling PSA architectures across diverse physics domains.

2. Coupling of Physics Engines, Emulators, and Numerical Backends

PSA systems tightly couple agentic orchestration layers to high-fidelity physics simulation codes. For inertial fusion capsule design, MADA uses MARBL, a radiation-hydrodynamics code, solving coupled PDEs—including mass conservation,

$$\frac{\partial \rho}{\partial t} + \nabla\!\cdot(\rho \mathbf{u}) = 0$$

momentum conservation,

$$\rho \frac{D\mathbf{u}}{Dt} = -\nabla p + \mathbf{S}_{\rm rad}$$

material energy evolution, and radiation energy transport. These are discretized on curvilinear (ALE) grids with matrix-free implicit radiation diffusion.

Emulator agents, typically PyTorch-based DCGAN surrogates, predict full-field outputs such as $r$–$t$ density plots, scalar traces, and thermodynamic profiles at orders-of-magnitude less computational cost than direct simulation, enabling rapid in-context evaluation and visual reasoning (Shachar et al., 2 Oct 2025). Emulators serve as an “experience bank”, allowing agents to recall and reason about the physics space without incurring costly reruns.

In agent pipelines focused on theoretical and computational physics (as in PhysMaster), symbolic and numerical computations are orchestrated within isolated execution sandboxes (Python/Julia) and critically coupled to retrieval augmented generation (RAG) of literature and prior knowledge (Miao et al., 22 Dec 2025).

3. Inverse Design, Optimization, and Meta-Reasoning

PSA agents execute inverse design and optimization loops in two principal modes:

1. Tool-driven: invoking optimizers (e.g., differential evolution, genetic algorithms) on formal physics-constrained objective functions, such as

$$\min_{x}f(x) = -[T_{\rm hs}(x)\,\rho R_{\rm hs}(x)]$$

subject to manufacturability and summation constraints.

2. Visual-feedback-driven: interpreting emulator-generated plots and trajectories (e.g., $T$–$\rho R$ curves) to guide sampling and refinement, particularly via visual cues such as crossing the Meldner ignition curve (Shachar et al., 2 Oct 2025).

Agentic meta-optimization exploits in-context updates: LLMs can use visual feedback and out-of-band results for gradient-like proposal refinements, resulting in rapid design space exploration and convergence.

A plausible implication is that PSA’s hybrid optimization approach—combining global (Latin hypercube or genetic) sampling, local sweeps, and visual meta-reasoning—can be generalized across domains, provided agent APIs and surrogate modeling are appropriately adapted.

4. Multimodal Explanation and Automated Evaluation

Recent PSA implementations extend beyond text to generate rich multimodal explanations, notably high-quality video solutions and visual proofs using Manim animation. The architecture comprises a solver agent for initial physics solution (chain-of-thought, formulas, visualization hints), a PlannerAgent storyboarder, CodingAgent for code synthesis, Manim Engine for rendering, and a vision-language module (VLM) for automated critique and layout refinement (Thole et al., 19 Jan 2026).

The evaluation pipeline computes 15 quantitative metrics (layout, readability, rendering accuracy, scene-content alignment, animation smoothness, symbol recognition, synchrony, error penalty, overall solution quality) and employs automated scoring formulas such as: $$\mathrm{OS} = 0.05\,S_Q + 0.10\,E_Q + 0.60\,\left(\frac{LQ+TR+ER+OSI+SCA}{5}\right) + 0.25\,(1 - \tfrac{\text{errors}}{\text{max\_errors}})$$

Empirical results for PSA’s video generation on 32 problems report 100% video completion, average OS of 3.8/5, and consistency across difficulty level and problem type. Visual-feedback-based refinement (driven by screenshot analysis and VLM critique) yields measurable improvements in scene-content alignment and visual polish.

However, systematic limitations remain, including layout inconsistencies, code hallucination (incorrect Manim calls), lack of dynamic feedback (single screenshot per scene), and delayed response for complex problems—signaling the need for multi-shot visual assessment and domain-aware RAG.

5. Physics-Informed Neural Approaches to PDE Solution

Extending PSA capabilities to PDE simulation and forecasting, frameworks such as PhysicsSolver integrate physics-informed neural networks (PINNs) and transformer attention modules. The architecture merges feed-forward (FNN) PINN backbones with an Interpolation Pseudo Sequential Grids Generator (IPSGG) and Halton quasi-Monte Carlo sampling for sparse supervision (Zhu et al., 26 Feb 2025).

PINNs encode the underlying physics via automatic differentiation and loss aggregation over residual, boundary, and initial constraints: $$\mathcal L_{\rm physics\_res},\quad \mathcal L_{\rm physics\_bc},\quad \mathcal L_{\rm physics\_ic}$$ with total loss

$$\mathcal L_{\rm total} = \lambda_1\,\mathcal L_{\rm physics} + \lambda_2\,\mathcal L_{\rm data}$$

Forecasting outside the training interval is enabled by temporal “attention” over sequences assembled by IPSGG, with single-step extrapolation yielding robust performance, particularly in domains where conventional PINNs and finite-difference methods fail.

Benchmark studies report PhysicsSolver achieves 1–2 orders of magnitude lower error than PINNs on convection, reaction, heat, and 2D Navier–Stokes equations—both for forward simulation and out-of-grid prediction.

6. Demonstrated Performance, Limitations, and Future Directions

Empirical demonstrations for PSA and related systems highlight rapid convergence and practical viability:

• MADA achieves simulated ignition in fusion capsule design, with hot-spot $T_{\rm peak} \sim 6.5~\mathrm{keV}$, areal density $\rho R_{\rm peak} \sim 0.3~\mathrm{g/cm}^2$, and gain $>1.2$ within five global feedback iterations (Shachar et al., 2 Oct 2025).
• PhysMaster compresses research timescales (weeks–months to hours), executes fully autonomous solution loops, and discovers new theoretical results independently (Miao et al., 22 Dec 2025).
• PSA generates instructive physics explanation videos with completion rates and quality scores surpassing baseline expectations; residual visual and logical errors persist, predominantly in video encoding and RAG conceptual accuracy (Thole et al., 19 Jan 2026).
• PhysicsSolver demonstrates robust PINN-transformer synergy, maintaining predictive accuracy in challenging PDE regimes and extrapolation tasks (Zhu et al., 26 Feb 2025).

Limitations include:

• Agent reasoning bottlenecks in ultra-abstract physics (e.g., string theory).
• Dependence on RAG and quality of external knowledge bases.
• Restricted error-correction and visual feedback loops, particularly in multimodal settings.
• Latency penalties for complex, interactive tasks.

Recommended future directions involve symbolic theorem prover integration, anomaly detection, richer visual feedback (multi-frame video critique over static screenshots), domain-knowledge RAG, and networked agent collaboration—including experimental apparatus integration (Thole et al., 19 Jan 2026, Miao et al., 22 Dec 2025).

7. Implications for Scientific Workflows and Educational Systems

The PSA paradigm bridges LLM-driven text reasoning, high-fidelity simulation, surrogate-model emulation, and multimodal (video/visual) explanation, supporting scalable, interpretable, and autonomous workflows in physics. PSA methodologies enable human-in-the-loop verification, pedagogical enhancement, and research acceleration, and are readily extensible to domains where agent APIs, code bases, and surrogate modeling can be modularly swapped (Shachar et al., 2 Oct 2025).

A plausible implication is that continued refinement of agent modularity, error-correction, and multimodal understanding will be necessary to achieve reliable, robust, and domain-general PSA systems, attuned to both scientific discovery and scalable education.