A modular view of recent deep inverse modelsfor inverse design problems

Abstract: A variety of deep learning models have been proposed in recent years for solving inverse design problems, termed deepinverse models (DIMs). In this work we show that modern DIMs can be abstracted as composed of three distinct modules: initialization,filtering, and local refinement. Work in just the last year has directly compared the performance of DIMs on benchmark problems however,based upon our insights, it is more appropriate to compare individual modules instead. In this work we compare the effectiveness andcharacteristics of the modules (filtering and local fine tuning) utilized in several modern DIMs, using four recently-proposed benchmarkinverse problems. Based upon our results, we find that a combination of modules from different existing DIMs, yielding a new overall DIMmodel, usually provides the best overall performance on the four benchmark tasks. Moreover, proving both empirically and theoratically,we show a unintuitive point that using a forward model to filter the inverse solutionsalwaysimproves or maintain the same averageone-shot performance, regardless of the accuracy of the forward model.

Recommended citation: Ren, S., Padilla, W., Malof, J., (2021). A modular view of recent deep inverse modelsfor inverse design problems.