AMLP-Analysis: Automated MLIP Validation

• AMLP-Analysis is a post-training validation module that rigorously evaluates machine learning interatomic potentials using DFT benchmarks, structural optimizations, and MD simulations.
• It leverages the ASE framework and automation via a simple YAML configuration to perform single-point calculations, geometry relaxations, and MD using protocols like NVE and NVT.
• Validation metrics such as MAE, RMSD, and RDFs confirm the high fidelity of ML-derived force fields compared to quantum-mechanical data.

Automated Machine Learning Pipeline Analysis (AMLP-Analysis) refers to the post-training and validation module within the Automated Machine Learning Pipeline (AMLP) for interatomic potential development and molecular simulation (Lahouari et al., 25 Sep 2025). It is designed to bridge ML-based force field training and downstream atomistic simulation tasks. Coupled with Python’s Atomic Simulation Environment (ASE), AMLP-Analysis enables comprehensive, automated evaluation of machine learning interatomic potentials (MLIPs) such as those based on the MACE architecture, providing end-to-end reproducibility and efficiency in modern atomistic simulations.

1. System Overview and Purpose

AMLP-Analysis serves as the key post-training component in the AMLP workflow, facilitating the transition from a fitted MLIP (e.g., a model trained on quantum-mechanical data) to practical validation and application in molecular simulations. Its role is to rigorously verify that the MLIP can match density functional theory (DFT) benchmarks in real-world tasks, including single-point property evaluation, structural (geometry and cell) optimization, and molecular dynamics (MD) simulations under realistic thermodynamic ensembles. The system is accessible via a simple, structured configuration file (config.yaml), enabling non-expert users to specify detailed simulation protocols without manual coding.

Integrated with the upstream AMLP modules, AMLP-Analysis ensures that datasets, input conversions, and simulation parameters are consistently managed throughout the model development pipeline.

2. Technical Architecture and Simulation Capabilities

AMLP-Analysis is constructed around the ASE framework. Its key simulation functionalities include:

• Single-point calculations: Direct evaluation of energies and forces for individual or batches of atomic structures.
• Geometry and cell optimizations: Structural relaxation using multiple optimizers (BFGS, LBFGS, FIRE, etc.) with user-set thresholds (e.g., |f|max for maximum force).
• Molecular dynamics (MD): Both microcanonical (NVE) and canonical (NVT) ensemble simulations:
  • NVE MD employs the velocity–Verlet algorithm. Energy conservation is quantified by the cumulative energy drift $$\Delta E = \frac{1}{N_\text{step}} \sum_k \frac{|E_k - E(0)|}{|E(0)|}$$ where $E(0)$ is post-equilibration baseline energy.
  • NVT MD supports both deterministic Nosé–Hoover chains and stochastic Langevin thermostats for temperature regulation and energy dissipation.
• Radial distribution functions (RDFs): Calculation of $g(r)$ for specific atom pairs or global structure, via

$$g(r) = \frac{1}{4\pi r^2 \rho N} \left\langle \sum_{i=1}^N \sum_{j\ne i} \delta(r - |r_i - r_j|)\right\rangle$$

providing insights on liquid structure, short- and long-range order, and the physical fidelity of the trained MLIP under simulation conditions.

Analysis outputs are standardized and compatible with ASE’s ecosystem, promoting reproducible research and downstream data handling.

3. Integration of LLM Agents

A distinctive feature of the broader AMLP system is the use of LLM agents in the initial stages of MLIP model development, preceding AMLP-Analysis. These agents:

• Parse user descriptions and atomic structure input to recommend quantum electronic-structure codes (e.g., VASP, CP2K, Gaussian).
• Specify code parameters (basis set, functional, dispersion method) for input file generation.
• Automate the workflow from quantum geometry optimization and AIMD input preparation through conversion to training-ready data formats (e.g., .json → HDF5).

Although AMLP-Analysis operates on the output MLIP and simulation dataset, its tight coupling with LLM-based upstream modules ensures that all subsequent validation is performed on models created with best-practice electronic-structure inputs, maximizing the accuracy and tractability of the ML potential.

4. Model Validation Metrics and Empirical Outcomes

The paper presents direct validation of AMLP-Analysis using the acridine polymorph system:

• Accuracy: MLIP trained within AMLP achieves mean absolute errors (MAE) of approximately 1.7–2 meV/atom for energies and 7 meV/Å for forces when compared to DFT calculations.
• Structural Reproduction: DFT-relaxed geometries are recovered to sub-Ångström accuracy (mean RMSD ≈ 0.048 Å).
• MD Stability: NVE MD simulations show robust energy conservation, with reported cumulative energy drift ~10⁻⁴.
• Physical Consistency: MLIP-generated RDFs for key atom pairs (e.g., C–N, N–N) across multiple temperatures match the qualitative and quantitative structure observed in DFT-based dynamics.

These results demonstrate that the AMLP + AMLP-Analysis combination can produce force fields that are not only broadly accurate in static properties but also stable and predictive in time-dependent MD applications, including thermal ensembles.

5. User Workflow and Automation

AMLP-Analysis is driven by a single YAML configuration, which encodes the sequence of analyses, simulation parameters, and convergence criteria. This high degree of automation relieves users from scripting or detailed workflow management, lowering barriers for non-experts. The model ensures that MLIP validation is documented, repeatable, and aligned directly with quantum reference data.

Table: Key Features of AMLP-Analysis

Capability	Implementation	Metrics Produced
Geometry optimization	ASE optimizers	RMSD, energy, forces
Molecular dynamics (NVE/NVT)	ASE, thermostats	Energy drift, RDF
Validation interface	config.yaml, ASE	MAE, RMSD, g(r)

6. Implications and Extensibility

The automation and modularity of AMLP-Analysis enable:

• Lowered entry barriers: Non-experts can rigorously validate MLIPs without manual workflow management.
• High-throughput and reproducible benchmarking: Automated calculation of standard metrics (e.g., energy drift, RMSD, RDF) enables systematic model comparison and sharing.
• Scalable integration: While exemplified on MACE, AMLP-Analysis is generalizable to other state-of-the-art MLIP frameworks, including NequIP, TorchMD, and FeNNol, due to its abstraction via ASE.

This framework facilitates robust, consistent validation of MLIPs, promoting the reliable extension of atomistic simulations to new chemical systems or larger simulation sizes.

7. Applications and Future Prospects

AMLP-Analysis is targeted at material scientists and computational chemists developing ML-derived force fields for molecular or materials modeling. Its demonstrated capabilities—sub-Ångström accuracy, robust MD stability, high-fidelity structure reproduction—make it suitable for research extending from small molecules to complex condensed-phase systems.

Anticipated future directions include broadening the pipeline’s support for additional MLIP methods, increased performance benchmarking across chemically diverse systems, and further enhancement of the LLM-guided workflow for automated input parameterization and out-of-the-box best practice recommendations.

In summary, AMLP-Analysis operationalizes rigorous MLIP validation, closes the loop from quantum data to simulation, and provides the necessary framework for credible, automated atomistic modeling workflows (Lahouari et al., 25 Sep 2025).