Figure 3 . The proposed synergy of multiscale and machine learning aspires to (i) accelerate the prediction of large-scale computational models, (ii) discover interpretable models from irregular and heterogeneous data of variable fidelity, and (iii) guide the judicious acquisition of new information towards elucidating the emergence of function in biological systems.
Here one can leverage ongoing developments in ML to accelerate the prediction of large-scale computational models. As a viable path forward, ML workflow can be implemented in three steps (see Figure 3) [4]: (1) Train deep neural network (NN) encoders that connect properties of the MSM model at scale 1 (e.g., MD or MC) to those at scale 2 (e.g., continuum model), using explicit MSM computations overextended parameter sets to cover all possible conditions; (2) implement the encoders in place of the coupling algorithms to bridge scales 1 & 2; (3) ensure that the properties profiles obtained from the two scales match by defining a cost-function that constrains the training of NNs in (1). The NN-based coupling of scales is expected to be robust, computationally efficient for MSM algorithms.
One challenge is to discover interpretable models from heterogeneous data of variable fidelity, and guide the judicious acquisition of new information towards elucidating the emergence of function in biological systems. This challenge can be addressed by subjecting the entire MSM model to contemporary data science and statistical methodologies, i.e., [142] sensitivity [1], evolvability [2], and robustness [143] analyses, uncertainty quantification [144], multi-fidelity modeling [145], and pattern discovery and model reduction [146].