Your issue may be addressed by adopting such a so-called (Multivariate) "Regression ANN" solution. Follow:
Asiltürk et al., "Modeling and prediction of surface roughness in turning operations using artificial neural network and multiple regression method", 2011 - http://booksc.org/book/3550770/15981c
Zain et al., " Regression and ANN models for estimating minimum value of machining performance ", 2012 - https://core.ac.uk/download/pdf/82063404.pdf
Prefer modern ANN implementations referred as "Geometric Deep Learning". Their multiresolution designs are really computationally efficient. In particular, multiresolution spline-wavelets might be a powerful tool. Follow:
Bronstein et al., " Geometric deep learning: going beyond Euclidean data ", 2017 - https://arxiv.org/pdf/1611.08097.pdf
Puy et al., " Unifying local and non-local signal processing with graph CNNs ", 2017 - https://arxiv.org/pdf/1702.07759.pdf
You may have a look at DeePMD-kit which provides an interface between deep learning representations of energy (based on Tensorflow) and MD simulations (based on LAMMPS). Follow:
[Section 6.2] Adie et al., " Deep Learning for Computational Science and Engineering ", 2018 - http://on-demand.gputechconf.com/gtc/2018/presentation/S8242-Yang-Juntao-paper.pdf
Wang et al., " DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamics ", 2017 - https://arxiv.org/pdf/1712.03641.pdf
https://github.com/deepmodeling/deepmd-kit
Some additional references to link deep learning based on multiresolution wavelets to MD simulations:
S. Mallat - " Understanding deep convolutional networks ", 2016 - https://www.di.ens.fr/~mallat/papiers/RSTA2015Published.pdf
Rinderspacher et al., " Wavelet Analysis for Molecular Dynamics ", 2015 - http://www.arl.army.mil/arlreports/2015/ARL-MR-0891.pdf
J.-A. Stende, " Constructing high-dimensional neural network potentials for molecular dynamics ", Master of Science Dissertation, 2017 - https://www.duo.uio.no/bitstream/handle/10852/59442/master.pdf?sequence=1