Filippo Guerranti
Nicholas GaoHi all,
I'm a PhD student in #MachineLearning at the technical university of Munich #TUM. I'm currently working on machine learning on graphs and machine learning-driven computional chemistry.
#ml #GraphNeuralNetworks #GNNs #compchem
rastinza@nicholasgao Hey Nicholas, what are you making with these Machine learning of chemical graphs?
Is it some kind of qsar?
Is it some kind of qsar?
@rastinza
Until now, my work has mostly been focused on machine learning potentials or ab-initio quantum chemistry, e.g.,
https://openreview.net/forum?id=apv504XsysP
Until now, my work has mostly been focused on machine learning potentials or ab-initio quantum chemistry, e.g.,
https://openreview.net/forum?id=apv504XsysP
@nicholasgao The idea looks very cool! Does it work?
@rastinza In our work, we found that by training a neural wave function on multiple geometries we do not sacrifice any performance but only have to train a single model reducing training times significantly.
While there is still work to be done to scale neural wave functions, we believe this is an important step in reducing their computational cost.
@nicholasgao Sorry, I did not read the article at all.
What do you mean by not sacrificing performance? You get results comparable to ab initio calculations? What order of approximation are we talking about?
What do you mean by not sacrificing performance? You get results comparable to ab initio calculations? What order of approximation are we talking about?
@rastinza we in fact perform highly accurate ab-initio calculations. In many small system, such ML-driven ab-initio methods report the lowest variational results in the literature. What we mean with "not sacrificing performance" is that compared to neural wave function-based baselines we don't lose any accuracy despite solving many Schrödinger equations simultaneously.
The plot below shows our Potential Energy Surface Network (PESNet) in comparison to other neural wave function-based methods.
@nicholasgao Wow, that's really impressive! Such small errors are not what I was expecting.
Can this be applied to larger systems?
Can this be applied to larger systems?
@rastinza In a recent preprint (https://arxiv.org/abs/2205.14962), we have shown that one can also get obtain continuous energy surfaces (denotes as PlaNet) for multi-dimensional data and "larger" molecules such as ethanol. Unfortunately, for systems larger than 40 electrons, these methods get quite expensive. So, their scaling in terms of size is still very much active research.
Sampling-free Inference for Ab-Initio Potential Energy Surface Networks
Recently, it has been shown that neural networks not only approximate the ground-state wave functions of a single molecular system well but can also generalize to multiple geometries. While such generalization significantly speeds up training, each energy evaluation still requires Monte Carlo integration which limits the evaluation to a few geometries. In this work, we address the inference shortcomings by proposing the Potential learning from ab-initio Networks (PlaNet) framework, in which we simultaneously train a surrogate model in addition to the neural wave function. At inference time, the surrogate avoids expensive Monte-Carlo integration by directly estimating the energy, accelerating the process from hours to milliseconds. In this way, we can accurately model high-resolution multi-dimensional energy surfaces for larger systems that previously were unobtainable via neural wave functions. Finally, we explore an additional inductive bias by introducing physically-motivated restricted neural wave function models. We implement such a function with several additional improvements in the new PESNet++ model. In our experimental evaluation, PlaNet accelerates inference by 7 orders of magnitude for larger molecules like ethanol while preserving accuracy. Compared to previous energy surface networks, PESNet++ reduces energy errors by up to 74%.