Multitask learning based on Hermitian operators

A non-transitory computer-readable medium having program code recorded thereon for multitask learning based on Hermitian operators, the program code being executed by a processor and comprising:

• program code to train a multitask machine learning (MTML) deep convolutional neural network (DCN) model to map an input formula/structure of a material onto a complex wave function state vector;
• program code to enforce a mathematical structure to machine learning of multiple properties of the material during the training to provide a trained MTML DCN model;
• program code to infer, by the trained, MTML DCN model, observable property matrices (OP) for each observable property of the material;
• program code to convert the observable property matrices into complex Hermitian operators for each observable property obs of the material as an expectation value (E) of each of the obs (E[obs]);
• program code to predict, by the trained MTML DCN model, target properties of the material, comprising a bandgap, a lattice constant, an elastic property, and/or a formation energy, according to the expectation value E of each observable property of the material E[obs], an observable parameters matrix (Mop), and the complex wave function state vector (PSI) by E[obs]=psi*.Mop.PSI as predicted target properties;
• program code to perform a loss function to reduce a difference between the predicted target properties (Ypred) and the target properties (Y) of the material to a threshold value;
• and
• program code to identify the material as a new material discovery when a unique combination of the predicted target properties is observed in the material according to the program code to perform the loss function.

A system for multitask learning based on Hermitian operators, the system comprising:

• a multitask machine learning (MTML) deep convolutional neural network (DCN) model, trained to map an input formula/structure of a material onto a complex wave function state vector, to enforce a mathematical structure to machine learning of multiple properties of the material during the training to provide a trained MTML DCN model, the trained, MTML DCN model to infer observable property matrices for each observable property (obs) of the material, to convert the observable property matrices OP into complex Hermitian operators for each observable property obs of the material as an expectation value (E) of each of the obs (E[obs]), to predict, by the trained MTML DCN model, target properties of the material, comprising a bandgap, a lattice constant, an elastic property, and/or a formation energy, according to the expectation value E of each observable property of the material E[obs], an observable parameters matrix (Mop), and the complex wave function state vector (PSI) by E[obs]=psi*.Mop.PSI as predicted target properties, to perform a loss function to reduce a difference between the predicted target properties (Ypred) and the target properties (Y) of the material to a threshold value, and to identify the material as a new material discovery when a unique combination of the predicted target properties is observed in the material according to the performance of the loss function.