Neural network potentials (NNPs), also called machine-learned interatomic potentials (or MLIPs), are machine-learned models that approximate solution of the Schrödinger equation, allowing scientists to run simulations with unprecedented performance, speed, and scale.
Conventional simulation techniques built on technologies like molecular mechanics (MM) and quantum mechanics (QM) force a tradeoff between speed and accuracy: MM-based simulations are extremely fast but suffer from known inaccuracies, while QM-based simulations rigorously approach experimental accuracy but are too slow to study long timescales or systems larger than a few hundred atoms.
NNPs are new technologies generated by training ML models to reproduce the results of QM simulations—the resulting models approach QM accuracy while running thousands or millions of times faster. Importantly, NNPs behave just like the conventional physics-based simulation methods they're trained on—so the exact same validated and trusted scientific workflows and applications built for legacy methods can be used with NNPs, all for a fraction of the cost.
NNPs allow us to obtain high-fidelity results at a fraction of the time and cost of conventional methods, which promises to transform molecular simulation across drug discovery and material science. Here's what one recent review has to say about the potential impact of NNPs:
The promise of NNPs lies in their potential to transcend the long-standing trade-off between accuracy and efficiency in simulations. Specifically, they offer the possibility of achieving quantum mechanics (QM)-level accuracy at the computational cost of classical MD… [this advance] opens the possibility of simulating large-scale complex systems starting solely from Schrödinger's equation. We can envisage a future where, through a simple web interface, we input a system of atoms and obtain a highly accurate depiction of their behavior at any desired scale. This would be a more general version of the way [AlphaFold] now enables the rapid prediction of the folded structure of a protein with a simple web interface.
Another review focused on materials science calls NNP progress “breathtaking” and concludes:
Given the relatively early stages of [NNP] technology and its undeniable successes, we can expect that this field will be very active for many years to come… [NNPs] are poised to become an indispensable part in the arsenal of methods available to the materials modeling community.
And a third review focused on water and aqueous systems argues that NNPs will enable previously impossible breakthroughs:
In summary, modern MLPs have created new opportunities for the investigation aqueous systems that would have been unimaginable with conventional methods for the foreseeable future… the remarkable progress made to date suggests that we can expect exciting new developments and some surprising breakthroughs in the years to come.
Rowan helps companies use NNPs to design and simulate molecules and materials faster and more accurately. We talk to researchers and keep up with this fast-moving area to understand which NNPs are appropriate for each task, and run our own rigorous benchmarking studies to get the most accurate data about NNP performance and accuracy.
Rowan makes it simple to run NNP-based simulations through our web-based computational chemistry platform. We've integrated leading NNPs like AIMNet2 and OMat24 into our platform, so that you can run geometry optimizations, transition-state calculations, frequencies, dihedral scans, and more with these methods—and automatically use the right computing hardware for the task. We've also built a custom pKa prediction workflow that uses NNPs for massive speedups so that you can get your results in minutes, not days.
Our internal researchers are experts in training specialized NNPs, evaluating their performance, and building high-performance computational workflows for specific tasks. If you or your company want to use the power of NNPs to accelerate research in your organization or work with us to develop your own fine-tuned models, contact us for an intro call!