Masgent: AI-Assisted Materials Simulation Agent

• Masgent is an AI-assisted materials simulation platform that integrates structure manipulation, VASP input creation, standardized DFT workflows, and machine learning simulations for rapid materials research.
• It employs natural-language processing to orchestrate complex simulation tasks, reducing the need for manual scripting and high-performance computing expertise.
• The platform standardizes simulation protocols and integrates machine learning potentials, achieving significant speedups and accurate benchmark results.

Masgent is an AI-assisted materials simulation agent that unifies structure manipulation, automated VASP input generation, DFT workflow construction and analysis, fast machine-learning potential (MLP) simulations, and lightweight machine learning utilities within a single platform. Designed for researchers in computational materials science, Masgent leverages LLMs to provide natural-language interaction, automating complex simulation tasks that traditionally require extensive scripting and high-performance computing expertise. By standardizing protocols and integrating simulation and data-driven tools, Masgent accelerates hypothesis testing, pre-screening, and exploratory research for both novice and expert practitioners (Liu et al., 28 Dec 2025).

1. Purpose and Scope

Masgent operates as an integrated, AI-driven framework for streamlining the full pipeline of materials simulations. The platform supports:

• Structure manipulation: Supercell creation, defects (vacancies, substitutions, interstitials), slab and interface construction, and SQS (Special Quasirandom Structures) generation.
• Automated VASP input generation: Produces INCAR, KPOINTS, POTCAR, and HPC job scripts based on best-practice, community-validated templates.
• DFT workflow construction and analysis: Implements standard workflows for convergence testing (ENCUT, k-points), equation of state (EOS) fitting, elasticity calculations, AIMD (ab initio molecular dynamics), and NEB (nudged elastic band) for migration barriers.
• MLP-based simulation: Unified API for SevenNet, CHGNet, Orb-v3, and MatterSim models, covering single-point calculations, EOS and elasticity fitting, and molecular dynamics.
• Lightweight ML utilities: Feature analysis (correlations, principal component analysis), data augmentation (CVAE), hyperparameter tuning (Optuna/TPE), and standard model evaluation metrics (RMSE, $R^2$).

Through a Pydantic-validated tool-calling system, Masgent enables natural-language queries such as “Compute the vacancy formation energy in LaCoO₃ with CHGNet”, automatically orchestrating retrieval/generation, input setup, workflow dispatch, and results summarization.

2. System Architecture

Masgent consists of three principal layers:

Layer	Description	Key Technologies
AI Mode	Conversational natural-language interface powered by LLMs.	LLM, Pydantic schemas
CLI Mode	Deterministic, menu-driven, script-free interface using Bullet library.	Bullet
Core Utilities	Structure tools, VASP input generators, workflow builders, MLP engines, and ML utilities.	Pymatgen, Icet, Optuna

The AI mode parses researcher intent and uses strict schema validation for tool calls. Session memory (active structures, previous outputs) supports multi-turn interactions. CLI mode mirrors all AI-mode capabilities, maintaining consistency. Core utilities contain routines for structure manipulation (Icet for SQS, pymatgen-analysis-defects for defect generation), VASP input generation, workflow orchestration, MLP wrappers, and ML modules (PCA, CVAE, Optuna hyperparameter search).

All operations are logged for full reproducibility. Directory structures and naming conventions are enforced, and only physically meaningful parameters are accepted via schema validation.

3. Algorithms and Computational Methods

Masgent automates simulation workflows leveraging robust algorithms:

• Density Functional Theory (DFT): Based on plane-wave DFT as implemented in VASP, using the total energy functional

$$E[n] = T_s[n] + \int v_{\text{ext}}(r)n(r)dr + \frac{1}{2}\iint \frac{n(r)n(r')}{|r-r'|}drdr' + E_{xc}[n]$$

with common exchange-correlation functionals such as PBE.

• Equation of State (EOS) Fitting: Generates a series of volumes $\{V_i\}$ near equilibrium $V_0$ and fits energies $\{E_i\}$ to the Birch–Murnaghan equation:

$$E(V) = E_0 + \frac{9V_0B_0}{16} \left[ \left((V_0/V)^{2/3}-1\right)^3 B_0^{\prime} + \left((V_0/V)^{2/3}-1\right)^2 (6-4(V_0/V)^{2/3}) \right]$$

• Elastic Constants: Applies small strains $\epsilon_j$ and extracts stress responses $\sigma_i$ to determine the stiffness tensor $C_{ij}$:

$\sigma_i = \sum_j C_{ij} \epsilon_j$

• Machine-Learning Potentials: The total energy is expressed as a sum over atomic contributions:

$E = \sum_i E_i = \sum_i \mathrm{NN}(G_i)$

with $G_i$ as local descriptors. Training employs MSE losses on energy and optionally forces.

• CVAE for Data Augmentation:

$$L_{\text{CVAE}} = -\mathbb{E}_{q_\phi(z|x,c)}[\log p_\theta(x|z,c)] + \mathrm{KL}(q_\phi(z|x,c)\,\|\,p(z))$$

where conditioning descriptors $c$ and latent variables $z$ capture data augmentation.

• Hyperparameter Optimization: Bayesian search via Tree-Structured Parzen Estimator (Optuna/TPE) to minimize objectives (e.g., RMSE).

Workflows are constructed programmatically via core utilities; all steps and outputs are organized in traceable directory trees.

4. User Interaction and End-to-End Automation

Masgent’s AI agent interprets complex natural-language queries, decomposes them into discrete steps, and executes end-to-end simulations. Example interaction:

1. Intent: “Perform a convergence test on ENCUT for bulk Al, then fit an EOS for the relaxed structure, and compare DFT versus CHGNet energies.”
2. Masgent flow:
   • Fetch Al structure from Materials Project
   • Run ENCUT tests (300–700 eV), collect the converged structure
   • Generate EOS inputs across scaled volumes
   • Launch static DFT calculations, fit EOS via Birch–Murnaghan
   • Repeat EOS energies via CHGNet (MLP)
3. Summarization: Reports convergence thresholds, equilibrium volumes, energy deviations (DFT vs MLP), and produces organized results with directory snapshots and analysis plots.

All parameter selection and workflow orchestration are automated, reducing setup times from hours to seconds.

5. Benchmarks and Representative Case Studies

Empirical benchmarks demonstrate substantial acceleration and reliability:

• Workflow setup time: Preparation for ten diverse materials (Al, Si, GaAs, MgO, NiO, TiO₂, MoS₂, LaCoO₃, La₂NiO₄, Cu) was completed in <30 seconds using Masgent’s AI mode, versus 1–3 hours manually.
• MLP accuracy and speed: Mean energy errors $|E_{\text{MLP}} - E_{\text{DFT}}| \lesssim 100$ meV/atom; speedups of $10^3$–$10^4$× for supercells of 8 to 1024 atoms, with MatterSim fastest.
• Case studies:
  • ENCUT convergence for Al: $\Delta E < 1$ meV/atom at ENCUT ≈ 400 eV.
  • EOS fitting for La₂NiO₄: $V_0 = 97.07$ ų.
  • NEB for La₂NiO₄ oxygen migration: barrier $A_e = 1082.7$ meV.
  • Elastic constants for Cu: $C_{11} = 216.6$ GPa, $C_{12} = 151.7$ GPa, $C_{44} = 105.9$ GPa.
  • ML model (Al–Co–Cr–Fe–Ni formation enthalpy): final $\text{RMSE}_{\text{train}} \approx 1.48$ eV/atom, $\text{RMSE}_{\text{val}} \approx 4.98$ eV/atom, $R^2 > 0.89$.

These results indicate reliable agreement between automated and manual workflows and demonstrate effective use of ML model acceleration within domain coverage.

6. Protocol Standardization, Democratization, and Best Practices

Masgent enforces strong protocol standardization for VASP inputs using community-validated templates (Materials Project sets), directory structure, and naming conventions for reproducibility. Only physically meaningful, strictly validated parameters are accepted via Pydantic schemas.

Democratization arises from the natural-language interface, which reduces scripting and HPC expertise requirements and supports broad accessibility for students and non-experts. Shared session logs and structured workflows enhance collaboration and knowledge transfer.

Recommended best practices:

• Confirm key AI-agent assumptions, particularly for convergence criteria and defect specification.
• Validate MLP predictions against DFT reference points before large-scale screening.
• Select MLP engines appropriate to material domain; user-supplied potentials recommended for out-of-domain cases.
• Employ rigorous session tracking (transcripts, directory snapshots) for reproducibility.
• Integrate Masgent-generated scripts with production workflow managers (e.g., FireWorks) and implement error-handling for HPC scaling.

7. Impact and Future Directions

Masgent represents a significant advancement in computational materials science by integrating structure manipulation, DFT automation, MLP acceleration, and lightweight ML tools within an AI-native environment. Standardized protocols and accessible interfaces promote reproducibility, rapid hypothesis testing, and reduce entry barriers for both new and experienced researchers. Plausible implications include further extensions to phonon workflows, automated job execution, and expanding the supported MLP/ML models, enhancing Masgent’s role as a next-generation materials simulation assistant (Liu et al., 28 Dec 2025).