Finally got around to reading the latest Orb paper introducing their new v3 model. The biggest takeaway is that we're starting to get into the territory of being able to simulate conditions we'd never been able to before due to how incredibly cost prohibitive they would be with traditional DFT. Paper here:
The authors introduce Orb-v3, the next generation of the Orb family of universal interatomic potentials. Models in this family expand the performance-speed-memory Pareto frontier, offering near SoTA performance across a range of evaluations with a ≥ 10× reduction in latency and ≥ 8× reduction in memory. Their experiments systematically traverse this frontier, charting the trade-off induced by roto-equivariance, conservatism and graph sparsity. Contrary to recent literature, they find that non-equivariant, non-conservative architectures can accurately model physical properties, including those which require higher-order derivatives of the potential energy surface.
It's inspired me to try some larger scale simulations. Everyone has access to high-end GPUs easily, and pretty cheap on Modal:
A100 40GB for $2.10/hour
A100 80GB for $2.50/hour
H100 80GB for $3.95/hour
H200 141GB for $4.54/hour
B200 192GB for $6.25/hour
Orbital researchers used a H200 for most of their testing.
With this accessibility and the relative simplicity of setting up MD simulations in ASE, there's really no reason not to give it a try.
I'm going to be focusing on metallic glass formation, crystallization, and annealing and see what I can learn there.
The breakthrough was in non-conservative architectures that keep a low memory and forward pass time, while still maintaining usable performance.
The Pareto frontier for a range of universal Machine Learning Interatomic Potentials. The 𝐾𝑆𝑅𝑀𝐸 metric assesses a model’s ability to predict thermal conductivity via the Wigner formulation of heat transport and requires accurate geometry optimizations as well as second and third order derivatives of the PES (computed via finite differences). The y-axis measure a model’s forward passes per second on a dense periodic system of 1000 atoms, disregarding graph construction time, measured on a NVIDIA H200. Point sizes represent max GPU memory usage. Y-axis jitter (+/- 5 steps/second) has been applied to allow visualization of overlapping points. Model families include a range of specific models with broadly the same architecture, but may be different sizes or trained on different datasets.
What's exciting me the most is being able to simulate systems where we can start to see emergent phenomena. One of the moments that really got me excited about materials science was Tim Duignan's (an Orbital Materials researcher) experiment showing a crystal nucleating and growing. https://x.com/TimothyDuignan/status/1797960944175427629
"Far more exciting is the possibility of applying Orb-v3 to study systems that have previously been impossible to simulate accurately due to the large number of atoms involved and the lack of existing accurately parameterized empirical forcefields [24]. Orb-v3 opens a new frontier where quantum mechanical accuracy can be maintained while exploring emergent phenomena arising from the collective behavior of thousands of atoms, such as crystal nucleation and growth [46], self-assembly of complex nanostructures such as metal organic frameworks [39], or phase diagrams of complex alloys [35]."
This sort of emergent behavior is all over materials science and chemistry. We've likely only scratched the surface of what's out there to discover.