Simulon: An AI-Assisted, PyTorch-Native Framework of Molecular Dynamics and Modeling

Area of Science:

• Computational chemistry and physics
• Artificial intelligence in scientific research
• Molecular dynamics simulations

Background:

• Modern molecular simulations face increasing complexity, necessitating integrated frameworks.
• Existing platforms often lack seamless integration of physical modeling, machine learning, and AI interaction.
• The need for efficient and user-friendly tools to handle complex molecular data is paramount.

Purpose of the Study:

• To introduce Simulon, an open-source, PyTorch-native molecular dynamics platform.
• To bridge the gap between molecular dynamics and artificial intelligence through a novel tensor-based architecture.
• To enable end-to-end learning over molecular data and accelerate simulations using GPU acceleration.

Main Methods:

• Developed a tensor-based architecture representing atomic systems, forces, and trajectories as differentiable tensors.
• Integrated a retrieval-augmented large language model (LLM) agent for natural language control of simulations.
• Implemented GPU acceleration and PyTorch software for computational efficiency.
• Ensured compatibility with classical and machine-learning potentials, validated against established engines like LAMMPS.

Main Results:

• Achieved substantial speedup in simulations through an optimized tensor-based computational kernel.
• Enabled conversational configuration, execution, and interpretation of molecular dynamics simulations via the LLM agent.
• Demonstrated quantitative agreement with established simulation engines while maintaining machine learning compatibility.
• Established a new paradigm for AI-assisted molecular modeling, accelerating scientific discovery.

Conclusions:

• Simulon offers a unified platform for AI-assisted molecular modeling, combining differentiable simulation, scalable computation, and natural language control.
• The platform streamlines complex simulation workflows, freeing scientists for more creative tasks.
• This integration accelerates scientific discovery by converging chemical information, data-driven potentials, and autonomous agents.