Technical Summary: Toward Exascale AI for Science – A Scalable AI Skill for Autonomous Microkinetics Discovery

Problem Statement

The discovery of complex material properties from first principles remains a grand challenge in science, particularly in the realm of reaction kinetics for complex material structures. While multiscale simulations combining quantum mechanical (QM) methods with surrogate foundation models represent the state-of-the-art, the process is often labor-intensive. Specifically, Nudged Elastic Band (NEB) calculations, used to determine minimum energy pathways (MEP) and activation energies, are highly sensitive to parameter choices and system configurations. This sensitivity frequently necessitates days or weeks of manual trial-and-error by domain experts. Furthermore, while Agentic AI promises to automate these workflows, robust frameworks capable of leveraging High-Performance Computing (HPC) resources and recovering from simulation failures remain in their infancy. Current Large Language Models (LLMs) often struggle with the "sycophantic" nature of generating incorrect scientific parameters, and there is a lack of one-size-fits-all solutions for autonomous scientific discovery.

Methodology

The authors propose a scalable, AI-driven framework centered on a specific "AI skill" called nebskill, designed to automate NEB calculations for microkinetics discovery. The methodology integrates three core pillars: agentic automation, materials-specific surrogate models, and large-scale HPC simulations.

1. Skill Design and Agentic Workflow

The nebskill is structured as a composable agent task that breaks the NEB process into five distinct steps:

1. Initialization: Environment setup, including downloading Machine Learning Interatomic Potential (MLIP) models and verifying reactant/product coordinates.
2. Relaxation: Relaxing reactant and product configurations.
3. Interpolation and Calculation: Interpolating between states and executing a two-phase calculation (standard NEB followed by Climbing-Image NEB).
4. Monitoring and Failure Recovery: If a calculation fails, the agent diagnoses the cause (e.g., "image collapse" or "kinking") and proposes new parameters (e.g., adjusting spring constants or the number of images) using a ReAct (Reasoning and Acting) approach.
5. Analysis and Reporting: Applying domain-expert methods to verify converged chemistry.

The agent utilizes a directory structure containing SKILL.md and INSTRUCTIONS.md files to manage context and prevent "context rot" by dividing tasks into subtasks. For decision-making, the agent interacts with an LLM inference endpoint (specifically the ALCF Inference Endpoint) via a system prompt to determine retry strategies, balancing the exploitation of accumulated knowledge with the exploration of new strategies.

2. Surrogate Models and Ground Truth

The study evaluates over ten State-of-the-Art (SOTA) universal MLIP models (including NequIP, MACE, and Allegro variants) as surrogates for graphite surface microkinetics. These equivariant networks encode atomic environments using spherical harmonics to ensure geometric consistency.

• Ground Truth: Quantum Molecular Dynamics (QMD) simulations were performed using the QXMD software with Density Functional Theory (DFT) as the ground truth. The system modeled CO2 sublimation from a graphite surface using projector-augmented wave (PAW) methods and GGA exchange-correlation functionals.

3. HPC Integration and Scalability

The framework was tested on the Perlmutter cluster at the National Energy Research Scientific Computing Center (NERSC) and the ALCF Inference Endpoint. The agent employs a two-dimensional parallelism strategy:

• Model Parallelism: Loosely coupled evaluation of multiple MLIP models (typically 5–10) distributed across different nodes.
• Image Parallelism: Tightly coupled distribution of NEB images (10–30) across GPUs within a single node to minimize communication latency for atomic force exchanges.

Key Results

1. Surrogate Model Performance

The study compared the activation energy profiles predicted by various pretrained MLIP models against DFT ground truth.

• Accuracy: The mace-mp-l model demonstrated the closest activation energy to the DFT result. The energy curves for nequip-oam and allegro-oam closely followed the DFT trajectory.
• Mechanism: All models successfully predicted the lactone-mediated CO2 sublimation mechanism, consistent with other QM calculations.

2. Scalability and Performance

Scalability tests on the Perlmutter cluster (HPE Cray system with NVIDIA A100 GPUs) demonstrated near-perfect parallel efficiency.

• Speedup: Using 16 GPUs (4 nodes), the system achieved a 3.75× speedup compared to the baseline (4 GPUs), reducing wallclock time from 116.9 minutes to 32.1 minutes.
• Efficiency: The distributed NEB computing approach effectively minimized communication overhead, validating the feasibility of scaling these workflows across leadership computing facilities.

3. Failure Recovery and Robustness

The agent successfully handled "resilient" jobs that failed on the first attempt.

• Failure Modes: Of 54 initially failed jobs that eventually converged, 51 suffered from "image collapse" (attributed to low spring constants), and 3 exhibited "kinking" (discontinuities in the energy profile).
• Recovery Strategy: The agent's default and most effective strategy was adjusting the spring constant. In complex cases (e.g., mace-mp), the agent iteratively adjusted parameters (e.g., increasing spring constant from 0.3 to 1.2, increasing NEB images, then decreasing the spring constant) to achieve convergence.
• Overhead: The communication cost with the LLM inference endpoint was low, averaging approximately 1,600 tokens per retry and 6–14 seconds of communication time, indicating high throughput.
• Uncertainty Quantification: By varying convergence criteria (maximum force between 0.1–0.3 eV/Å), the study quantified the sensitivity of different MLIP models. Models like equiformer-v3 and nequip-oam-m showed extremely low standard deviation, while others like matris-oam showed higher variance, aiding in the systematic selection of robust surrogates.

Significance and Claims

The paper claims that this work represents a significant step toward Exascale AI for Science by demonstrating how AI skills can transform complex domain workflows into robust, scalable capabilities.

• Autonomous Recovery: The study proves that AI agents can autonomously recover from failed simulations, reducing the need for expert intervention in labor-intensive tasks like NEB calculations.
• Systematic Evaluation: The framework enables the systematic evaluation of surrogate model reliability and the quantification of prediction uncertainty, which is critical for reproducible science.
• Scalable Architecture: By integrating agentic reasoning with HPC resources, the authors show that it is possible to manage simulations on national supercomputing infrastructure using open models, paving the way for "AI scientists" capable of executing grand challenge workflows.
• Modest Scope: The authors position the work not as a fully autonomous "one-size-fits-all" solution, but as a "co-scientist" approach that provides immediate, practical value for communities exploring agentic workflows. The goal is to encapsulate domain expertise into reusable AI skills that can operate across distributed computing facilities to accelerate materials discovery.