MINDES: Data Mining for Inverse Design
A typical approach in the design of new products is “direct design”, where we start with an initial specification of a product and adjust it until we get a product with the desired output characteristics. However, what is needed is the inverse approach, where, starting with the properties we want in the end product, we determine the inputs which will likely yield these properties.
This idea of inverse design is gaining popularity in a variety of domains ranging from design of wind turbines to semiconductor process and device modeling. The basic idea is to exploit the wealth of data generated by experiments and computer simulations to guide the process of design. This can be done either through the use of optimization techniques, such as simulated annealing and genetic algorithms, or through the application of ideas from data mining.
In this project, we will use techniques from statistics and data mining to:
- identify key inputs relevant to an output, and
- build surrogate models relating the outputs to the inputs.
This ARRA-funded SciDAC-e project, is a collaboration between the SDM SciDAC Center and the Center for Inverse Design EFRC . The EFRC is proposing a new approach to materials science where, through the use of inverse design techniques, they want to find the structure of a material given its properties.
Our work is being done in context of materials for solar cells using data from quantum mechanical simulations being run by the EFRC. If successful, this work will allow us to identify the inputs to the simulations intelligently and reduce the number of expensive simulations which must be run, thus leading to the discovery of new materials, not by chance, but by design. The ideas being developed in this project are also applicable to inverse design problems in other domains, such as the design of experiments in fusion, as well as inverse problems in general.
This is a collaboration with Drs. Mayeul d’Avezac and Alex Zunger at NREL.