News

Radical AI Unveils TorchSim, an AI-Powered Engine for Materials Simulation

• By John K. Waters
• 04/15/2025

Artificial intelligence research startup Radical AI has launched TorchSim, a next-generation PyTorch-native engine designed to dramatically accelerate atomistic simulations. The open-source software, designed for the Machine Learning Interatomic Potential (MLIP) era, offers computational speeds up to 100 times faster than traditional tools like ASE and 100 million times faster than conventional DFT methods, according to the company.

MLIPs are computational models that use machine learning techniques to approximate the potential energy surfaces of atomic systems with high accuracy and significantly lower computational cost compared to traditional quantum mechanical methods like Density Functional Theory (DFT).

The release of TorchSim marks Radical AI’s first public demonstration of a sweeping new approach to scientific research—one meant to replace large teams and slow, linear workflows with autonomous, AI-augmented processes driven by individual scientists. With TorchSim, the company aims to shift the foundation of materials discovery from legacy frameworks to GPU-accelerated, batched simulation models that can evaluate, optimize, and correlate materials properties in real-time.

"Scientific progress has followed the same playbook for over a century," the company said in a statement. "TorchSim is our first step toward replacing that with a self-driving, AI-powered model of science—one where a single researcher can solve multiple problems simultaneously."

TorchSim integrates a suite of high-performance features, including batched molecular dynamics and optimization algorithms—such as NVE, NVT, NPT, and FIRE—executed through a unified PyTorch-based interface. The platform’s architecture intelligently utilizes GPU memory, dynamically managing simulations across models like MACE and Fairchem with efficiency that reflects commercial-scale deployment needs.

At the heart of TorchSim is SimState, a tensor-based structure capable of representing and manipulating both single and multiple atomic systems. Supporting auto-batching, real-time trajectory compression, and compatibility with common machine learning libraries, TorchSim aims to become a foundational layer for AI-driven labs and next-gen research.

Radical AI emphasized its broader ambitions, framing TorchSim not as a standalone tool but as part of its proprietary "Materials Flywheel" ecosystem, a commercial engine intended to unlock rapid, scalable materials innovation for sectors such as aerospace, energy, defense, and semiconductors. While TorchSim is open source, Radical AI plans to commercialize the materials it helps discover.

The company is also calling for open-source contributions, requiring a signed Contributor License Agreement (CLA) for all participation. TorchSim remains in an experimental phase, and all submissions will undergo mandatory review by maintainers.

With TorchSim, Radical AI is staking a claim on the future of materials science—one where algorithms, not armies of researchers, drive discovery at unprecedented speed and scale.