💥I am happy to share our work on next gen transferable ML force fields for bio-organic / molecular chemistry. MACE-OFF23 is a series of pretrained MACE models accurately describing molecules in vacuum as well as condensed phase with a wide range of tests.

RT @DrJimFan: Hitchhiker's guide to rebranding:.- Machine learning -> statistical mechanics.- Loss function -> energy functional.- Optimize….

RT @chaitjo: There's been many new Geometric GNNs in the past couple weeks combining 3D equivariance and topological ideas (simplices, cell….

RT @PythonFZ: Check out the interactive demo on Zero Shot Molecular Generation via Similarity Kernels for our prepr….

RT @karpathy: Have you ever wanted to train LLMs in pure C without 245MB of PyTorch and 107MB of cPython? No? Well now you can! With llm.c:….

RT @ask1729: 1/ Introducing a new method, the Gaunt Tensor Product, to improve the efficiency of E(3)-equivariant neural networks and speed….

Check out this great summary of our foundation model for chemistry / materials science by @lars__schaaf . This thread hints at the emergent behaviour of our model with applicability well beyond the training domain of crystals. 👇👇👇.

And to try the models quickly just install mace from pip install mace-torch. And import them as an ASE calculator. calc = mace_off(model="medium", device='cuda').

As well as tagging all other co-authors who were integral to this project! @IlyesBatatia Nick Browning @venkatkapil24 @w_c_witt @ColeGroupNCL @joshhorton93 @MagdauIoan and Gabor of course!.

This paper is the final large piece in my PhD journey! . I have to thank to all my co-authors and co-workers without whom it wouldn't have been possible! .In particular @jhmchem co-first author and drug discovery expert.

Finally and crucially the MACE-OFF23 models are already fast, capable of simulating 3-4 million MD steps / day (3-4 ns with 1 fs timestep). But we will have a new release very shortly with MUCH faster custom CUDA implementation 🚀🚀

We are of course not limited to small peptides and can simulate larger proteins in explicit solvent such as crambyn with vibrational spectra showing accurate secondary structure (without any extra training like GEMS)

MACE-OFF23 can also simulate peptides in explicit solvent to obtain free energies or to simulate the folding dynamics of larger peptides such as Ala15. Below is free energy surface of Ala3 in explicit solvent vs Amber

Water (the biology of physicists) can also be simulated using MACE-OFF23. We get a reasonable description (both spectra and RDF-s). But water is not all, we also simulated 109 different small molecule liquids in NPT to obtain densities roughly matching experimental ones.

Next we tried the model for molecular crystals computing Raman spectra (with MACE fitted polarisabilities) including nuclear quantum effects and enthalpies of sublimation both closely matching experimental values.

We validated the models on small molecule torsions and found to almost match the accuracy of the reference method compared to CCSD(T) on 600+ unseen molecules

We present 3 models of increasing accuracy parameterised for 10 (H C N O P S F Cl Br I) chemical elements trained to reproduce accurate (hybrid DFT with large basis) QM energies and forces. The models are available at

RT @venkatkapil24: If you are around in London don't forget to swing by @KingsCollegeLon tomorrow afternoon for the @tyc_london new starter….

RT @jrib_: Great news! Following hot on the heels of the MACE-MP-0 foundation model that hit arXiv ( 10 days ago, t….

RT @maurice_weiler: MACE is my favorite among recent E(3)-equivariant MPNN models. Instead of relying only on the usual two-body messages,….