Large Language Model Based Agent for Automated Discovery in Computational Physics

The paper introduces PhyNex, an autonomous agent that leverages large language models and domain-specific tools to systematically explore and optimize scientific solutions in computational physics, successfully matching or exceeding human-designed state-of-the-art methods across diverse tasks such as dielectric spectrum prediction, Max-Cut heuristics, and quantum battery optimization.

The Gist

Imagine you are a master chef trying to invent a new recipe. You know exactly what the dish should taste like (the goal) and you have a list of allowed ingredients and kitchen rules (the physical constraints). However, you don't know the exact amounts of spices or the precise cooking times. Traditionally, you would have to spend months or years tasting, adjusting, failing, and tweaking your recipe until it's perfect.

This paper introduces PhyNex, a new kind of "robot sous-chef" designed to do this tasting and tweaking for you, specifically for problems in computational physics.

Here is how PhyNex works, using simple analogies:

1. The Robot Chef's Strategy

Instead of guessing wildly, PhyNex acts like a very organized, persistent tinkerer.

• The "One-Step-at-a-Time" Rule: Imagine you have a complex machine. Instead of rebuilding the whole thing from scratch, PhyNex changes just one small part at a time (like swapping a gear or tightening one screw). It then tests the machine.
• The Scorecard: Every time it makes a change, it gets a score. If the score goes up, it keeps that change. If the score goes down, it tries something else.
• The "Lesson Book": This is the robot's superpower. If a change causes the machine to break (a "bug"), PhyNex doesn't just give up. It writes down why it broke and how to fix it in a shared "Lesson Book." If another robot branch tries to make the same mistake later, it checks the book and avoids the error. This means the more it tries, the smarter it gets.

2. The Three Challenges (The "Recipes")

The authors tested PhyNex on three very different scientific "recipes" to see if it could outperform human experts:

• Challenge A: Predicting Light (The Crystal Prism)

  • The Task: Scientists have crystals and want to know exactly how they will interact with light (like a prism splitting light into colors). Usually, this requires expensive, slow computer simulations.
  • The Result: PhyNex figured out a way to predict these light patterns directly from the crystal's shape. It discovered a specific rule: "Light absorption must always be a positive number" (you can't have negative light). By adding this simple rule, it became more accurate than the human-designed models.

• Challenge B: Cutting the Graph (The Party Split)

  • The Task: Imagine a party where people are connected by friendships (a graph). You want to split the guests into two groups so that the maximum number of friendships are "cut" (people in different groups). This is a classic math puzzle.
  • The Result: PhyNex invented a new strategy for handling "popular" people (hubs) who know everyone. It decided to make decisions about these popular people first. This approach was much better at splitting the group than the methods humans had previously designed.

• Challenge C: Charging the Quantum Battery (The Energy Sprint)

  • The Task: Quantum batteries are tiny, futuristic batteries that can charge incredibly fast, but they are chaotic and hard to control. Scientists need to find the perfect "charging schedule" to get the most energy out without the battery exploding or losing energy.
  • The Result: PhyNex found two different ways to charge the battery. One way was a smooth, steady rhythm (like a calm heartbeat), and another was a cautious strategy that prepared for the worst-case scenarios. Both methods extracted more energy than the human-designed methods, especially in the early stages of charging.

3. Why This Matters

The paper claims that PhyNex can solve these problems in about 12 hours, a task that might take human researchers months of trial and error.

• It's Transparent: Unlike some AI that is a "black box" (you don't know how it works), PhyNex leaves a trail of breadcrumbs. You can look at its "Lesson Book" and see exactly which small change made the biggest improvement.
• The Division of Labor: The paper suggests a new way for science to work:
  • Humans define the rules, the goals, and the physical laws (the "What" and "Why").
  • PhyNex handles the boring, repetitive work of trying thousands of combinations to find the best solution (the "How").

In short, PhyNex is an automated explorer that navigates the vast landscape of scientific solutions, learning from its own mistakes and finding better paths than humans can find alone, all while keeping a clear record of how it got there.

Technical Summary

Technical Summary: PhyNex – An LLM-Based Agent for Automated Discovery in Computational Physics

Problem Statement

Scientific discovery in computational physics often involves optimizing quantitatively evaluable objectives subject to physical constraints. While researchers excel at formulating these problems, the process of iteratively refining methods, debugging implementations, and tuning solution strategies is labor-intensive, often requiring months or years. Existing automated approaches face significant limitations: modular neural-symbolic architectures often lack generalizability, and evolutionary program-search methods, while flexible, obscure the causal link between specific code modifications and performance gains. Furthermore, many autonomous research agents are tailored to specific task classes, making adaptation to new domains costly.

There is a need for a system that can:

1. Generalize across diverse computational physics problems.
2. Provide interpretable attribution of performance improvements to specific algorithmic components.
3. Navigate the search space of executable programs without relying on gradient-based optimization (since the mapping from code to score is non-differentiable).

Methodology: The PhyNex Framework

The authors introduce PhyNex, an autonomous agent designed to systematically explore the solution space of scorable scientific tasks. PhyNex couples Large Language Model (LLM)-guided search with domain-specific computational tools that enforce physical consistency.

Core Architecture

The framework operates as a closed-loop agent (Fig. 1) defined by the following components:

• Problem Formulation: A task $T$ is defined as $(X, Y, U)$, where $X$ is the input space, $Y$ is the output space, and $U$ is a set of domain-specific tools (simulators, data loaders, evaluators) provided by the scientist. The goal is to find an executable program $\omega$ that maximizes a scoring function $M(\omega)$.
• Progressive Local Search: PhyNex does not perform global restructuring. Instead, it refines a solution through localized, single-component modifications. At each step, the LLM proposes a targeted change $\Delta\omega$ to a parent program. This ensures that changes in the score can be directly attributed to specific algorithmic choices.
• Knowledge Accumulation: The system maintains a global knowledge base ($K_{global}$) of "lessons" derived from both successful and failed attempts.
  • Rectification: If a candidate program fails (runtime error), the error and diagnostic output are fed back to the LLM to generate a fix.
  • Failure Lessons: Successful repairs generate lessons describing the failure mode and the solution. These are added to $K_{global}$ to prevent redundant failures in subsequent branches.
• Depth-Guided Parallel Exploration: PhyNex launches $K$ independent search trees in parallel, each starting from a different initial solution.
  • Tree Logic: A branch continues only if a modification improves the score; otherwise, it terminates.
  • Coupling: All trees share the $K_{global}$ knowledge base, allowing a failure encountered in one trajectory to be avoided in others.
  • Trajectory Logging: Every modification is logged with its score change, creating an explicit, interpretable exploration trajectory.

Key Contributions

1. Autonomous Algorithmic Discovery: PhyNex autonomously identifies solutions that match or exceed state-of-the-art (SOTA) human-designed baselines across three distinct domains without requiring extensive prompt engineering.
2. Interpretability and Attribution: By restricting modifications to single components and logging the resulting score changes, PhyNex produces exploration trajectories that reveal which design choices drive performance. This allows researchers to understand the causal mechanisms behind improvements (e.g., identifying that a specific activation function or scheduling strategy was the primary driver of success).
3. Physical Consistency via Tooling: The framework enforces physical constraints not through the LLM's internal knowledge alone, but through the tool set $U$ (e.g., simulators, evaluators), ensuring all candidate solutions operate within valid physical regimes.

Experimental Results

PhyNex was validated on three representative problems, achieving search-averaged improvements over human baselines within 12 hours of computation.

Task 1: Spectral Prediction of Semiconductors

• Objective: Predict frequency-dependent dielectric spectra from crystal structures.
• Baseline: Human-designed graph neural network (GNN) from Ref. [22].
• PhyNex Performance: Achieved search-averaged similarity coefficients (SC) exceeding the baseline.
  • $\text{Im}(\bar{\epsilon}_{100})$: $0.810 \pm 0.011$ (vs. $0.78$ baseline).
  • $\text{Re}(\bar{n}_{300})$: $0.951 \pm 0.003$ (vs. $0.94$ baseline).
• Key Insight: PhyNex autonomously introduced physically motivated constraints, such as Softplus activation to enforce non-negative optical absorption and baseline offsets for refractive indices, which were identified as the primary drivers of improvement.

Task 2: Probabilistic-Circuit Max-Cut Optimization

• Objective: Design variational algorithms for Max-Cut on regular and Barabási–Albert (BA) scale-free graphs.
• Baseline: R-PAOA [23].
• PhyNex Performance:
  • Regular Graphs: Improved normalized mean cut from $0.649$ to $0.743$ (2-regular) and $0.567$ to $0.652$ (3-regular) using only 4 parameters (vs. hundreds in R-PAOA).
  • BA Graphs: Improved normalized mean cut from $0.561$ to $0.603$.
• Key Insight: The agent discovered degree-aware gate scheduling (prioritizing hub nodes) and temporal correlations between gates, effectively exploiting the heterogeneous structure of scale-free networks.

Task 3: Charging-Protocol Optimization for Dicke Quantum Batteries

• Objective: Optimize time-dependent control protocols to maximize ergotropy in the chaotic coupling regime.
• Baseline: Human-designed Soft Actor-Critic (SAC) approach.
• PhyNex Performance:
  • Guided Exploration (SAC prior): Achieved a $7.78\%$ improvement at the 80k training checkpoint.
  • Open Exploration (No prior): Achieved a $5.90\%$ search-averaged improvement at the 80k checkpoint and slightly outperformed the baseline at 480k steps.
• Key Insight: The agent identified that replacing prioritized experience replay with uniform sampling and adding a smoothness penalty to the actor loss improved performance by reducing overfitting to quantum noise and preventing destabilizing control fluctuations.

Significance and Claims

The paper claims that PhyNex demonstrates a practical division of labor in scientific research:

• Scientists define the objectives, constraints, and evaluation metrics (via the tool set $U$).
• Automated Systems navigate the methodological search space, handling the trial-and-error loop of implementation and hyperparameter tuning.

The authors emphasize that PhyNex does not replace physical insight but accelerates the path from problem specification to effective implementation. The system's ability to produce interpretable exploration trajectories is highlighted as a critical feature, allowing researchers to extract mechanistic insights (cause-effect patterns) that can inform future algorithm design. The work suggests that for problems with scorable objectives and moderate evaluation costs, systematic exploration driven by LLMs can substantially accelerate scientific discovery.

Limitations Noted by Authors:

• The framework is limited to tasks with scorable objectives and moderate per-evaluation costs (excluding prohibitively expensive simulations like large-scale DFT).
• Exploration is local; it may miss qualitatively different solution families requiring non-local jumps.
• The search objective in Task 3 was defined at a specific checkpoint (80k), and results are most pronounced there.