/Data-driven multi-fidelity surrogate modeling for accelerated materials discovery

Data-driven multi-fidelity surrogate modeling for accelerated materials discovery

Gent | About a week ago

Combine approximations and costly experiments into smart and fast multi-dimensional models to accelerate the discovery of tomorrow's semiconductor materials.
The discovery of new functional materials is increasingly limited by the cost of generating reliable data: accurate simulations and experimental characterization are too expensive to exhaustively explore both a vast chemical space of candidate compositions and the multiple, often competing, properties a material must satisfy at once. Cheaper approximate sources of information exist, but current multi-fidelity surrogate models are ill-suited to exploit them. They typically assume a small number of ordered fidelities, a simple and constant relation between them, and comparable, evenly sampled data at each level. In practice, the available data are heterogeneous, sparse and unevenly distributed across fidelities, affected by different noise and biases, sometimes contradictory, and related in ways that vary across the design space. This PhD project develops data-driven surrogates that learn these relations from the data themselves, so that every expensive observation is exploited to the fullest and low-cost data can fill the gaps around it, with quantified uncertainty. On top of these surrogates, the project designs active learning strategies that decide which candidates to evaluate, at which fidelity, and toward which trade-off among the competing objectives, subject to real-life constraints on budget, time, and availability of resources, and validates the methodology on materials screening problems formulated with domain experts.
 


 

Required background: Computer Science/Engineering, Data Science

Type of work: 70% modeling/simulation, 20% experimental, 10% literature

Supervisor: Ivo Couckuyt (UGent)

Co-supervisor: Sebastian Rojas Gonzalez

Daily advisor: Stefan Lefever

The reference code for this position is 2027-110. Mention this reference code in your application.

Who we are
Accept marketing-cookies to view this content.
Cookie settings
imec inside out
Accept marketing-cookies to view this content.
Cookie settings

Send this job to your email