Eamon Whalen and Caitlin Mueller
ASME. J. Mech. Des. February 2022; 144(2): 021704
Surrogate models have several uses in engineering design, including speeding up optimizations, interpolating measurements, portability, and protecting intellectual property. Traditionally, surrogate models require that all training data conform to the same parametrization (e.g. design variables), which limits design freedom and prohibits the reuse of historical data.
This paper proposes Graph-based Surrogate Models (GSMs) for trusses, which can accurately predict a trusses’ structural performance given only its geometry and loads as inputs. Since the GSM does not rely on design variables to make predictions, the model can be trained and used on designs from multiple sources.
This paper then explores transfer learning within the context of engineering design, and demonstrates how re-training a GSM that was initially trained for another task requires significantly less data than training a GSM from scratch. Transfer learning is effective even when the data sets are of different topologies, complexities, loads or applications. The result is a more flexible and data-efficient surrogate model for truss design.
Left: A previously trained Graph-based Surrogate Model (GSM) can be re-trained on a new data set with differing geometry, loads or topology. Right: Pre-training significantly increases the data efficiency of the GSM. In these results, a pre-trained GSM trained on 20 designs (N=20) outperforms a fresh GSM trained on 500.
