(224a) Data-Efficient Probabilistic Model Learning with Embedded High-Fidelity Knowledge for Biomanufacturing in Deep Space Manned Missions

The NASA Center for the Utilization of Biological Engineering in Space (CUBES) leverages systems biology and systems engineering to realize life support for long, deep space manned missions on Mars. The aim of the center is to advance the practicality of multi-function, multi-organism biomanufacturing for production of pharmaceuticals, cell-based treatments/therapeutics, and biopolymers for on-demand 3D printing using in situ Martian resources [1-2]. Robust predictive modeling and control capabilities for the biomanufacturing system are crucial for decision-support and mission optimization under strict safety-critical requirements for life support on Mars [3].

In this talk, we will discuss how the abundancy of data generated from high-fidelity simulations, offline experiments, and real-time process operation can be systematically leveraged for learning probabilistic data-driven models for biomanufacturing systems. In particular, we will present a Bayesian learning framework that embeds high-fidelity knowledge and uncertainty quantification for data-efficient probabilistic model learning for a Mars life support biomanufacturing system. The proposed data-driven model learning framework is comprised of deep neural network surrogates for first-principles bioprocess models, trained based on simulations, combined with a Bayesian neural network with variational layers that acts as a real-time “corrector” to model hard-to-model phenomena that cannot be sufficiently captured by the first-principles models. This probabilistic nature of the learning-based model, expressed through the uncertainty of the variational layers, allows for quantifying the probabilistic uncertainty of model predictions, making the models especially suitable for robust decision-making and optimization. Furthermore, embedding the first-principles knowledge in the probabilistic learning-based models enhances data efficiency in model learning, while the ability to adapt the models using real-time data improves their predictive capability by accounting for unmodeled system behaviors.

We will demonstrate the proposed model learning framework on an integrated system of bioprocesses for Martian biomanufacturing. The process consists of production of biological feedstocks through microbial processing of in situ resources to fix carbon and nitrogen from the atmosphere, followed by downstream production of food, pharmaceuticals, and biopolymers suitable for 3D printing. The integrated bioprocesses are modeled using an advanced first-principles modeling tool, echusOverlook, developed in CUBES. We will discuss how the probabilistic learning-based models with embedded high-fidelity knowledge can enable evaluating different Mars mission scenarios and life support technologies in terms of the equivalent system mass metric, as well as reliability and cost-benefit considerations, under uncertainties in models and data.

[1] Menezes A.A., Montague M.G. , Cumbers J., Hogan J.A. , Arkin A.P., “Grand challenges in space synthetic biology,” J. R. Soc. Interface, vol. 12, no. 113, p. 20150803, 2015.

[2] Menezes A.A., Montague M.G. , Cumbers J., Hogan J.A. , Arkin A.P., “Towards synthetic biological approaches to resource utilization on space missions,” J. R. Soc. Interface, vol. 12, no. 102, p. 20140715 2015.

[3] Makrygiorgos, G., Gupta, S. S., Menezes, A., & Mesbah, A. “Fast Probabilistic Uncertainty Quantification and Sensitivity Analysis of a Mars Life Support System Model,” IFAC-PapersOnLine, In Press, 2020.

[4] Liang, F. “Bayesian neural networks for nonlinear time series forecasting,” Statistics and Computing, vol. 15,13-29, 2005.