🏂🏸Adversarial attacks to refine molecular energy predictions
Researchers at MIT have found a new quantitative estimate of the uncertainty of molecular energies using neural networks. Neural networks are often used to predict new resources, speeds, and capabilities orders of magnitude faster than traditional methods such as quo-mechanical simulation. The results obtained can be unreliable, since ML-models are interpolated, it is possible that they fail when applied to the operational data of an external dataset. This is especially for predicting the "potential energy" (PES) or energy map of a molecule in all its configurations. To solve these problems, scientists have proposed safe zones of a neural network using adversarial attacks. The actual simulation is performed only for small parts of the molecule, and the data is fed into the neural network, which learns to predict the same properties for the rest of the molecules. These methods have been successfully tested on new materials, including catalysts for the production of hydrogen from water, cheaper polymer electrolytes for electric vehicles, magnets, etc. However, the accuracy of neural networks depends on the correctness of training data, and incorrect predictions can have disastrous consequences.
One way to find out the uncertainty of a model is to run the same data through several versions of it. To do this, the researchers had several neural networks predicting a potential surface based on the same data. If the network is confident in the prediction, the difference between the outputs of different networks is minimal and the surfaces converge more. Otherwise, the predictions of the various models vary greatly, producing a series of outputs, any of which may be the correct surface.
Forecast scatter represents the uncertainty at a particular point. The ML-model should indicate not only the best forecast, but also the uncertainty of each of them. However, each simulation can take tens to thousands of CPU hours. And to get meaningful results, you need to run multiple models at a sufficient number of points.
Therefore, the new approach only selects data points with low forecast confidence. These molecules are then modified slightly to increase the uncertainty. Additional data is computed for these molecules through simulation, and then the original training pool is added. The neural networks are trained again, and a new set of uncertainties is calculated. This process is repeated until the uncertainty associated with various points on the surface becomes well defined and cannot be further reduced.
The proposed approach has been tested on zeolites - cavernous crystals, selective forms and use in catalysis, gas separation and ion exchange. Modeling large zeolite structures is very expensive, and the researchers show how their method can provide significant savings in computer simulations. But an adversarial approach to retraining neural networks increases performance without significant computational costs.
https://news.mit.edu/2021/using-adversarial-attacks-refine-molecular-energy-predictions-0901