A new physics-informed machine learning model can now simulate molecules at extreme temperatures for unprecedented lengths of time, finally solving a problem that has plagued computational chemists for years. Researchers at The University of Manchester have developed an AI system that keeps molecular simulations stable and accurate even when molecules are pushed to 1000 Kelvin (about 1,340 degrees Fahrenheit), opening doors to faster discoveries in drug development, new materials, and sustainable chemistry .

Why Do Molecular Simulations Usually Fall Apart at High Temperatures?

For decades, scientists have relied on machine-learned potentials (MLPs), which are AI models trained to approximate how quantum mechanics governs molecular behavior. These models work reasonably well at room temperature, but they have a critical weakness: when molecules experience heat, movement, or structural distortion, the simulations become unstable. Atoms start behaving unphysically, sometimes collapsing together or flying apart, making long-term simulations impossible .

This instability has forced researchers to either rely on expensive supercomputers running slower, more accurate quantum mechanical calculations, or accept the limitations of their AI models. The Manchester team, led by Professor Paul Popelier and including researchers Bienfait Kabuyaya Isamura, Olivia Aten, and Mohamadhosein Nosratjoo, decided to tackle this head-on by embedding deep physical knowledge directly into their AI architecture.

How Does the New AI Model Keep Molecules Stable?

The team built their model using Gaussian process regression, a statistical machine learning technique that learns patterns from data while maintaining uncertainty estimates. The key innovation was feeding the model detailed information about how atoms naturally interact based on quantum physics rules, giving the AI a realistic foundation for predicting molecular behavior .

But the real breakthrough came from a seemingly small mathematical choice: the "prior mean function." This function acts like a starting point or anchor for the model's predictions. By carefully selecting this function, the researchers discovered they could give the AI the correct "starting point" to create and sustain stable simulations even when molecules were stretched, heated, or shaken .

"For years, the community has focused on accuracy benchmarks, but we've shown that the real test is whether a model can survive the unpredictable situations molecules encounter during simulation. Our models don't just survive, they actively correct unphysical behavior," explained Professor Paul Popelier, Professor of Computational Chemistry at The University of Manchester.

Professor Paul Popelier, Professor of Computational Chemistry, The University of Manchester

Unlike conventional approaches that simply try to predict molecular structure, the new model uses real-world physical principles to prevent atoms from collapsing or flying apart when molecules enter high-energy states. This enables reliable simulations far beyond room temperature, a milestone rarely achieved by machine-learning force fields .

What Results Did the Researchers Achieve?

The team demonstrated their model's robustness through rigorous testing. They ran 50 independent simulations, each lasting 10 nanoseconds, totaling 0.5 microseconds of stable dynamics. This might sound brief, but it represents a major milestone: most machine-learning force fields cannot maintain stability for this duration at extreme temperatures .

The model proved its versatility by keeping even highly flexible molecules stable throughout testing. These included aspirin, serine, and glycine, molecules that are notoriously difficult to simulate because they have many degrees of freedom and can adopt multiple conformations. The model also demonstrated the ability to repair distorted structures and accurately reproduce known conformations, such as those of alanine dipeptide, a key benchmark molecule in computational chemistry .

"We discovered that simply shifting one mathematical function transforms model behavior entirely. With the right choice, the model consistently prevents molecular catastrophes and becomes extraordinarily robust," noted Bienfait Kabuyaya Isamura, PhD Candidate in the Department of Chemistry at The University of Manchester.

Bienfait Kabuyaya Isamura, PhD Candidate, Department of Chemistry, The University of Manchester

How to Apply This Technology to Real-World Research

• Drug Development: Researchers can now simulate how drug molecules behave under physiological conditions and at elevated temperatures, accelerating the identification of stable, effective compounds without relying on expensive supercomputers.
• Materials Discovery: Scientists can explore how new materials respond to extreme conditions, such as high temperatures in jet engines or thermal stress in industrial processes, enabling faster optimization of material properties.
• Computational Efficiency: The model runs on standard CPU hardware at speeds comparable to or faster than leading neural-network-based potentials that require high-end GPUs, making advanced molecular simulation accessible to more research groups.
• Biomolecular Systems: Long-timescale accuracy is now possible for studying protein folding, enzyme dynamics, and other biological processes that require extended simulation periods to capture meaningful behavior.

What Makes This Different From Previous Approaches?

The Manchester team's approach differs fundamentally from most AI models in computational chemistry. Rather than treating the AI as a black box that learns patterns from data alone, they integrated quantum mechanical principles directly into the model's architecture. This physics-informed approach means the model doesn't just memorize training data; it understands the underlying rules governing molecular behavior .

The computational efficiency is particularly noteworthy. Previous high-accuracy models often required specialized hardware like graphics processing units (GPUs) to run at reasonable speeds. The new model achieves comparable or better performance on standard central processing units (CPUs), democratizing access to advanced molecular simulation for research groups without access to expensive computing infrastructure .

What's Next for This Technology?

The research, published in Communications Chemistry, represents a proof-of-concept that opens multiple research directions. The team is now extending their approach to include electron correlation effects, which are important for accurately modeling certain types of chemical bonding and reactivity. They are also developing more transferable descriptors, which would allow the model to generalize better across different types of molecules and chemical systems .

The implications extend beyond academic research. The ability to reliably simulate molecules at extreme conditions could accelerate development of new catalysts for sustainable chemistry, high-temperature materials for aerospace applications, and more stable drug formulations. By reducing the need for extensive physical experiments during optimization, this technology could significantly shorten development timelines and reduce costs across the chemical and pharmaceutical industries.