Physics-enhanced machine learning to optimize a state-of-the-art manufacturing process

We are bringing together modern techniques in Bayesian machine learning to optimize process parameters for a state-of-the-art semiconductor fabrication process. The challenge is that while the process contains complex underlying dynamics and some degree of stochasticity, there are a large number of process parameters and a small number of samples to use for model training and validation. Thus, efficiently modelling the process phase space is of paramount importance. The goal is to predict several key material properties, and our approach involves embedding physics knowledge into the models so that the parameter space can be searched efficiently and in accordance with known physical constraints. The bayesian approach will enable us to estimate both the optimal parameters and also the sensitivities of the process to its inputs. These modules will be combined into a hierarchical machine learning platform to replace existing design of experiment requirements and eliminate calibration runs. The purpose of this work is to demonstrate that accurate models of manufacturing processes can be made without the use of computationally expensive and unnecessarily complex techniques. The expected benefit for the partner company is to ultimately productize this approach and deliver to manufacturers to increase throughput and yield.

Faculty Supervisor:

Rafael Kleiman

Student:

Partner:

Circuit Mind Inc

Discipline:

Engineering

Sector:

Professional, scientific and technical services

University:

McMaster University

Program: