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].