Boris Kozinsky's group at Harvard: Understanding dynamics of materials with computational physics + chemistry and machine learning.
Harvard University, Cambridge MA
Joined March 2019
- Tweets187
- Following486
- Followers2.1K
- Likes92
If you’re in the Boston/Cambridge area on Friday, December 5 (right after the Materials Research Society Fall Meeting), join us at Harvard University for our NequIP Tutorial.
👉 Please RSVP here to attend and receive event updates:
docs.google.com/forms/d/e/1F…
Excited to share that our NequIP and Allegro foundation potentials, trained by @Kavanagh_Sean_, are up on Matbench Discovery. Check them out at nequip.net/ ! 🚀
Materials Intelligence Research @ Harvard retweeted
Great to see our initial set of NequIP & Allegro foundation potentials released and on matbench-discovery!
Along w/excellent accuracies, we also find our model to (𝘤𝘶𝘳𝘳𝘦𝘯𝘵𝘭𝘺) be the fastest of leading foundation potentials – see our posters below 🏎️
Preprint incoming!
great news for all the people who over the past year reached out to ask why Nequip and Allegro were missing from Matbench Discovery. they're finally up as of today thanks to outstanding work by @Kavanagh_Sean_ and the MIR group @bkoz37.
Leaderboard: matbench-discovery.materials…
Materials Intelligence Research @ Harvard retweeted
Machine learning can be powerful for understanding defects, but currently sufficient only in select cases.
MLIPs (& geometric/electrostatic tools in doped) allow screening for challenging 'non-local' defect reconstructions (split vacancies) in all ICSD/MP solids, w/caveats 🔗
Materials Intelligence Research @ Harvard retweeted
In @NatureComms, we report how the sequence of preparing Pd8Au92/SiO2 catalysts tunes the size of Pd ensembles on the nanoparticle surface
@HarvardCCB @hseas @ucmerced @NorthwesternU @stonybrooku @Materials_Intel @ChristianReece @UUBeta @AizenbergLab
nature.com/articles/s41467-0…
Last month, we released a major update to the NequIP framework that fully leverages PyTorch 2.0 compilation for MLIPs. It’s significantly faster, easier to use, and more versatile than before.
Preprint: arxiv.org/abs/2504.16068
Code: github.com/mir-group/nequip
nequip.readthedocs.io
With @_MitKotak from @AtomArchitects, we also added custom GPU kernels for the Allegro tensor product. These combined improvements made Allegro 5-18x faster than before, and for large models, enabled simulations with 40-50 times more atoms than what was previously possible.
We look forward to adding more features. For example, we just released support for OpenEquivariance in NequIP: github.com/mir-group/nequip/….
(Thanks Vivek and Austin passionlab.github.io/OpenEqu…)
Feel free to reach out if you have questions or ideas: github.com/mir-group/nequip?…
Materials Intelligence Research @ Harvard retweeted
A machine learning framework that can predict with quantum-level accuracy how materials respond to electric fields, up to the scale of a million atoms.
bit.ly/3ZXmgzf
ALT Gray spheres and colored arrows that illustrate charge dynamics in barium titanate atoms.
Materials Intelligence Research @ Harvard retweeted
Excited to see our AI model for electric fields featured by Harvard @hseas ! 🚀
Link to paper 👉 nature.com/articles/s41467-0…
seas.harvard.edu/news/2025/0…
Our Allegro-Pol model extended the Allegro architecture to predict how materials respond to external electric fields while enforcing physical rules. It could describe vibrational, dielectric, and ferroelectric behavior for systems up to millions of atoms!
nature.com/articles/s41467-0…
We applied Allegro-Pol to study the temperature-dependent and frequency-dependent ferroelectric response of BaTiO3, revealing the underlying mechanisms of nucleation and growth that govern ferroelectric domain switching.
Materials Intelligence Research @ Harvard retweeted
Beyond happy to announce today Allegro-pol, a machine-learning framework that predicts how materials respond to electric fields with quantum-level accuracy, capturing vibrational, dielectric, and ferroelectric behavior at the million-atom scale! 🚀
nature.com/articles/s41467-0…
Discover our simple guidelines for training accurate and transferable equivariant ML interatomic potentials for ionic liquid mixtures. Test them on your systems and let us know your results! @JPhysChem #IonicLiquids #MachineLearning DOI: doi.org/10.1021/acs.jpclett.…
💡Open position for Professor in Applied Mathematics at Harvard @hseas with focus on Computing and AI for Science, Engineering, and Society. Emphasis is on development of applications with strong mathematical and computing foundations. Apply by 12/31/23. academicpositions.harvard.ed…
🏆Nice way to start the winter break⛄️
Beyond happy and very honored to receive my tenure promotion at Harvard @hseas. Most grateful to all my colleagues and collaborators, especially the amazing members of our @Materials_Intel group, whose work made this possible. Now the fun begins😀
NequIP architecture from @Materials_Intel + compute, data & brains @Google = another leap in accuracy, this time in the universe of all known materials.
When evaluated on @jrib_'s MatBenchDiscovery, the model again does well (note that we only eval'ed on this long after the potential was trained, we never explicitly tried to optimize for this).
NequIP potentials trained at scale @GoogleDeepMind: GNoME models discover 2.2M (380,000 stable) crystals, expanding the space of materials known to humanity (OQMD+MaterialsProject+WBM) by x10! Already 736 of these materials synthesized by LBNL and others.
dpmd.ai/GNoME-AI