TSAgent: An Agentic Workflow for Autonomous Transition State Search

The paper introduces TSAgent, an autonomous agentic workflow that automates transition state searches at the DFT level with accuracy comparable to human experts, successfully navigating complex, multi-step simulation challenges to reproduce established scientific scaling relationships.

The Gist

Imagine you are trying to find the highest point on a foggy, mountainous landscape to cross from one valley to another. In the world of chemistry, this "highest point" is called a Transition State. It's the exact moment a chemical bond breaks or forms, acting as the gatekeeper that determines how fast a reaction happens. Finding this spot is crucial for designing better catalysts (materials that speed up reactions), but it's incredibly difficult.

Traditionally, finding this spot is like trying to navigate that foggy mountain with a blindfold on, guided only by a map that updates once a day. It requires a human expert to run complex, expensive computer simulations, check the results, realize something went wrong, adjust the settings, and try again. This process is slow, manual, and prone to getting stuck.

Enter TSAgent, a new "digital explorer" described in this paper. Here is how it works, using simple analogies:

1. The Problem: The Foggy Mountain

Chemical reactions happen on a "Potential Energy Surface," which is like a 3D map of hills and valleys.

• Valleys are stable chemicals (reactants and products).
• The Hilltop is the Transition State.
• The Fog represents the complexity. The computer simulations (called DFT) are so heavy and slow that they take hours or days to run. When they fail, the error messages are often confusing, and the "map" might show a weird shape that only an expert can interpret.

2. The Solution: TSAgent (The Autonomous Hiker)

TSAgent is an AI system designed to climb this mountain entirely on its own. It doesn't just run a script; it acts like a persistent, thinking hiker who uses a specific loop: Plan → Execute → Analyze → Replan.

• The Planning Agent (The Strategist): This is the brain. It looks at the starting point (reactants) and the destination (products) and draws a route. It decides, "First, I'll relax the ground, then I'll try to climb the hill."
• The Execution Agent (The Climber): This is the worker. It sends the instructions to the supercomputer to run the actual simulation. It waits for the result (which might take a day) and then comes back.
• The Visual Analyzer (The Eye): This is the unique part. When the climb fails, the system doesn't just look at numbers (like "force is too high"). It actually looks at 3D pictures of the atoms, just like a human chemist would squinting at a screen. It asks, "Did the atoms crash into each other? Did a piece of the molecule fall off?"

3. How It Handles Failure (The "Oops" Moment)

In the past, if a simulation failed, a human had to step in. TSAgent handles this automatically:

• Scenario A: The computer says, "The force is stuck."
  • TSAgent's Reaction: "Ah, the hiker is taking steps that are too big and overshooting the path. I'll tell the hiker to take smaller steps."
• Scenario B: The computer says, "The path looks weird."
  • TSAgent's Reaction: "Wait, looking at the 3D picture, I see the reaction isn't one simple jump. It's actually two jumps with a rest stop in the middle. I need to split this mission into two separate climbs."

This ability to diagnose the specific type of failure and change the strategy on the fly is what makes it "agentic." It doesn't just follow a pre-written script; it adapts like a human expert would.

4. The Results: How Good Is It?

The authors tested TSAgent in three ways:

• The Obstacle Course: They gave it 100 different chemical reactions from a standard benchmark. TSAgent successfully found the transition state 83% of the time.
• The Human vs. Machine Race: They pitted TSAgent against three human experts (PhD chemists with years of experience) on 10 new, difficult reactions.
  • Success Rate: The humans succeeded 73% of the time (on average). TSAgent succeeded 70% of the time.
  • The Takeaway: TSAgent matched the performance of top human experts but did it without needing a human to sit there and click buttons. The humans spent about 47 minutes per successful case manually fixing errors; TSAgent did it autonomously.
• The Real-World Test: They asked TSAgent to reproduce a famous scientific rule (the Brønsted–Evans–Polanyi relationship) regarding ammonia dissociation. TSAgent successfully recreated the complex patterns found in the original study, proving it can handle real scientific investigations, not just textbook examples.

5. The Catch (Limitations)

The paper is honest about what TSAgent can't do yet:

• It needs a starting map: You still have to tell it where the reactants and products are. It can't invent the starting materials from scratch.
• It's expensive: While it saves human time, it still uses a lot of computer time (CPU hours). In fact, it used slightly more computer time than the humans on average because it sometimes tries a few different strategies before finding the right one.
• It's not magic: If the computer crashes or the physics is too weird, it might still get stuck, though it tries to recover.

Summary

TSAgent is a self-driving car for chemical discovery. Instead of a human driver manually steering, checking the GPS, and fixing the engine when it sputters, TSAgent is the car that drives itself, looks at the road, realizes a tire is flat, changes its route, and keeps going until it reaches the destination. It has proven it can drive as well as a professional human driver, opening the door to exploring thousands of chemical reactions that were previously too tedious to study.

Technical Summary

Technical Summary: TSAgent – An Agentic Workflow for Autonomous Transition State Search

1. Problem Statement

Identifying transition states (TSs) on potential energy surfaces (PES) is a fundamental bottleneck in the mechanistic study of catalytic materials. While Density Functional Theory (DFT) provides the necessary quantum-mechanical accuracy, locating a TS is not a single calculation but a complex, long-horizon workflow. This process involves a sequence of atomistic simulations (geometry optimizations, Nudged Elastic Band searches, and vibrational frequency analyses) characterized by:

• Delayed and Asynchronous Feedback: DFT calculations on high-performance computing (HPC) clusters often take hours to days, preventing real-time decision-making.
• Heterogeneous Failure Modes: Failures are not uniform; they range from convergence issues (e.g., forces plateauing) to qualitative physical errors (e.g., atom collisions, unphysical bond breaking, or the presence of stable intermediates indicating a multi-step reaction).
• Multimodal Diagnosis Requirements: Detecting these failures often requires a joint analysis of scalar convergence metrics (forces, energies) and visual inspection of 3D atomic geometries. Scalar criteria alone are insufficient to distinguish between a failed optimization and a physically valid but complex reaction path.
• Combinatorial Explosion: In complex catalytic networks (e.g., CO2 reduction to urea), the number of potential elementary steps and competing TSs grows exponentially, making manual, static scripting infeasible for large-scale mechanistic studies.

Existing automated pipelines often rely on static scripts or machine learning interatomic potentials (MLIPs) that lack the adaptive, closed-loop reasoning required to recover from DFT-level failures.

2. Methodology: TSAgent Workflow

The authors propose TSAgent, an agentic workflow designed to automate TS searches directly at the DFT level. The system operates on a persistent Plan–Execute–Analyze–Replan loop, decoupling high-level scientific reasoning from low-level operational execution.

2.1 Agent Architecture

The workflow utilizes a multi-agent system powered by GPT-5.4, differentiated by role-specific system prompts and toolsets:

• Planning Agent (PA): The "brain" of the system. It generates a SimulationPlan composed of typed SimulationStep objects (Geometry Optimization, NEB, Vibrational Frequency Analysis). Upon failure, the PA diagnoses the root cause by synthesizing numerical diagnostics and visual evidence to revise the strategy.
• Execution Agent (EA): Orchestrates the execution of the plan. It interacts with the simulation environment (VASP on SLURM-managed HPC) via specialized sub-agents:
  • Inputs Generator: Prepares VASP input files (INCAR, POSCAR, etc.).
  • Job Submission: Manages HPC job submission and monitoring.
  • Outputs Parser: Extracts structured scalar diagnostics (forces, energies, convergence flags) from VASP output files using pattern-based extractors to avoid context window overflow and hallucination.
  • Visual Analyzer: Renders orthogonal 2D projections of the reaction path (NEB images) to detect qualitative events like bond breaking, rotation, desorption, or atom collisions that scalar metrics miss.

2.2 Adaptive Replanning and Multimodal Diagnosis

The core innovation lies in the multimodal diagnosis loop. When a step fails or stalls:

1. Evidence Aggregation: The EA aggregates scalar diagnostics (e.g., force plateaus, lack of imaginary frequencies) and visual summaries (e.g., "stable intermediate detected," "atom overlap").
2. Diagnosis: The PA analyzes this evidence against expert-crafted debugging guidelines and its own domain reasoning.
3. Strategy Revision: The PA generates a revised plan. Interventions are not limited to parameter tuning (e.g., reducing MAXMOVE or POTIM) but can include structural changes to the workflow, such as:
   • Splitting a single TS search into two independent searches if a stable intermediate is detected (decomposing a multi-step reaction).
   • Cropping the reaction path around a local maximum to increase resolution.
   • Restarting with different convergence criteria or disabling the climbing-image algorithm temporarily.

The loop continues until a physically validated TS (a first-order saddle point with exactly one imaginary frequency) is found or the compute budget is exhausted.

3. Key Contributions

1. TSAgent Framework: Introduction of a closed-loop, agentic workflow that autonomously diagnoses and recovers from DFT-level TS search failures without human intervention. It mirrors the iterative debugging process of human experts.
2. Direct DFT-Level Automation: Unlike prior works that rely on MLIP approximations for TS search, TSAgent operates directly at the DFT level, ensuring the results are consistent with the high-fidelity theory used in benchmark construction.
3. Multimodal Reasoning: The system integrates visual analysis of atomic geometries with scalar convergence diagnostics, addressing failure modes that are invisible to purely numerical scripts.
4. Benchmarking Against Experts: A quantitative comparison demonstrating that the agent achieves success rates comparable to human domain experts.
5. Scientific Utility Demonstration: The successful reproduction of Brønsted–Evans–Polanyi (BEP) scaling relationships for NH3 dissociation, validating the agent's ability to handle real-world scientific investigations beyond curated datasets.

4. Experimental Results

4.1 OC20NEB Benchmark Evaluation

TSAgent was evaluated on a stratified subset of 100 reaction pathways from the OC20NEB heterogeneous catalysis benchmark (covering dissociation and transfer reactions).

• Success Rate: TSAgent achieved an 83% overall success rate (95% CI: 74.5–89.1%).
  • Dissociation reactions: 90% success.
  • Transfer reactions: 76% success.
• Impact of Replanning: The one-shot success rate (without replanning) was only 40%. Performance improved incrementally with each replan round, demonstrating that robust TS discovery relies on iterative, evidence-driven adaptation rather than a single initial strategy.
• Comparison to MLIP Baseline: TSAgent outperformed a baseline pipeline using the Universal Model for Atoms (UMA) foundation MLIP, highlighting the difficulty of the DFT regime compared to MLIP-based searches.

4.2 Comparison with Human Experts

The agent was benchmarked against three PhD-level DFT practitioners on 10 held-out OC20NEB reactions.

• Success Rate: TSAgent achieved a 70% success rate, compared to a human expert average of 73 ± 12%.
• Effort: Human experts spent an average of 47 minutes per successful case on manual debugging, structure inspection, and input modification. TSAgent required no such manual operator effort.
• Compute Cost: TSAgent consumed approximately 3,100 more CPU-hours per case than the human average. This is attributed to the agent's need for additional iterations to converge on a winning strategy, whereas experts can sometimes identify the dominant failure mode immediately.

4.3 Real-World Application: BEP Scaling

TSAgent was used to compute activation barriers and reaction energies for NH3 dissociation on 12 pure metal and single-atom alloy (SAA) surfaces.

• The agent successfully reproduced the linear BEP scaling relationships found in literature ($R^2 = 0.78$ for pure metals, $R^2 = 0.91$ for SAAs).
• It correctly identified the distinct scaling trends between pure metals and SAAs, demonstrating its capability to generate scientifically valid data for catalyst design.

5. Significance and Claims

The paper claims that TSAgent represents a significant step toward automating the "last mile" of computational mechanistic studies. By decoupling scientific reasoning from operational execution and closing the loop with multimodal diagnostics, the system addresses the specific challenges of long-horizon, asynchronous, and failure-prone DFT workflows.

The authors emphasize that TSAgent is not merely a parameter sweeper but a system capable of reasoning through failure. Its ability to decompose complex reactions and adapt strategies based on visual and numerical evidence allows it to tackle problems that static scripts cannot. While the system currently requires significant compute resources (a cost shared by human experts in terms of total CPU time, though with higher efficiency in human-hours), it offers a pathway to scale mechanistic studies to the combinatorial complexity of real-world catalytic networks, potentially accelerating the generation of training data for next-generation reactive potentials and generative models.

Limitations Acknowledged:

• The workflow assumes reasonable reactant and product geometries are provided; generating these remains a separate challenge.
• The evaluation is limited to 100 reactions due to the high cost of DFT calculations.
• Automated TS assignments, while rigorously validated by the agent, lack a physical guarantee and could propagate errors if deployed without oversight in downstream datasets.