University of Manchester Chemists Launch "Disaster-Proof" AI Model for Stable Molecular Simulations at Extreme Temperatures

University of Manchester researchers develop a robust AI model capable of stable molecular simulations at extreme temperatures without supercomputers.

By: AXL Media

Published: Mar 31, 2026, 6:10 AM EDT

Source: Information for this report was sourced from University of Manchester.

Overcoming the Instability of AI Molecular Potentials

Machine-learned potentials (MLPs) have become a staple in modern chemistry for approximating quantum mechanical behavior without the prohibitive cost of traditional supercomputing. However, these models have historically suffered from a critical "fragility" problem: when simulated molecules are subjected to high temperatures or significant structural distortions, the AI often loses its way, causing atoms to collapse together or fly apart. This instability has long prevented researchers from conducting the long-duration simulations necessary to understand how complex molecules behave in real-world, high-energy environments.

Integrating Deep Physics into Gaussian Processes

Led by Professor Paul Popelier, the Manchester team—including Bienfait Kabuyaya Isamura, Olivia Aten, and Mohamadhosein Nosratjoo—has solved this challenge by embedding "deep physical knowledge" directly into their AI architecture. Using a technique called Gaussian process regression, the researchers fed the model granular data on how atoms interact based on the fundamental laws of quantum physics. This provides the AI with a realistic "starting point," ensuring that even when a molecule is stretched or shaken, the model understands the physical boundaries it must operate within.

The "Prior Mean Function": A Mathematical Safety Net

The team discovered that a specific mathematical component, known as the "prior mean function," was the key to total model stability. By correctly configuring this function, the AI was granted a baseline understanding of atomic behavior that acts as a corrective force. PhD candidate Bienfait Kabuyaya Isamura noted that shifting this single function transformed the model’s behavior entirely. Instead of spiraling into "molecular catastrophes" when molecules entered high-energy states, the Manchester model demonstrated an unprecedented ability to self-correct and maintain structural integrity.