Teddy Koker

Currently @periodiclabs. ML PhD student at @MIT; AI for materials.

[Teddy Koker @teddykoker]

Introducing Nequix MP, a new foundation model for materials.

Example of using Nequix to simulate the dissolution of a NaCl nanocrystal in water:

1/n

[Teddy Koker @teddykoker]

Special thanks to collaborators @ganganabhijeet, @_MitKotak, Jaime Marian, and @tesssmidt!

n/n

[Teddy Koker @teddykoker]

New preprint! Phonon fine-tuning (PFT)

Phonons arise from small lattice vibrations and control thermodynamic properties.

Instead of hoping interatomic potentials get curvature right indirectly, we fine-tune directly on phonons, aligning Hessians with DFT force constants.

1/n

[Teddy Koker @teddykoker]

The method is model agnostic, enabling future universal interatomic potentials to achieve accurate vibrational properties. We hope this inspires more data generation and fine-tuning on other DFT-derived properties or even experiment observables.

7/n

[Mit @_MitKotak]

Nequix update! Training time is further down to 100 A100 hours and inference is 5x faster, thanks to OpenEquivariance kernels + PyTorch 2.0 backend.

1/n

[Teddy Koker @teddykoker]

@chuinwei_tan Thanks! Yes, Muon is only applied to 2d weights, which are all the weights in the models except the RMSNorm scale, where Adam is used.

[Teddy Koker @teddykoker]

Introducing Nequix MP, a new foundation model for materials.

Example of using Nequix to simulate the dissolution of a NaCl nanocrystal in water:

1/n

[Teddy Koker @teddykoker]

We hope Nequix serves as a strong, efficient baseline for future work on accessible materials foundation models!

n/n

[Teddy Koker @teddykoker]

allowing us to train our model in just 500 GPU hours, a fraction of the training cost of other competitive models. We also demonstrate a 10x improvement in inference speed over the leading model, eSEN (@xiangfu_ml et al.).

4/n

[Ameya Daigavane @BigAmeya]

Really excited to (finally) share the updated JAMUN preprint and codebase! We perform Langevin molecular dynamics in a smoothed space which allows us to take larger integrator steps. This requires learning a score function only at a single noise level, unlike diffusion models.

[Teddy Koker @teddykoker]

@evilmathkid @wgilpin0 @TexasScience 2. Throughout training I added an extra loss to ensure the final configuration is stable (to encourage the NCA to stop once it gets to the solution).

[Teddy Koker @teddykoker]

@evilmathkid @wgilpin0 @TexasScience It has been a while, and I didn't spend much time on this but I recall a few important things:
1. I used an iterative training process where I first try to get the NCA to solve in n=1 steps, n=2, up to some max number of steps. With each step I decreased the learning rate.

[Rianne van den Berg @vdbergrianne]

🚀 After two+ years of intense research, we’re thrilled to introduce Skala — a scalable deep learning density functional that hits chemical accuracy on atomization energies and matches hybrid-level accuracy on main group chemistry — all at the cost of semi-local DFT. ⚛️🔥🧪🧬