Physics-Based Machine Learning Modeling for Platform Design, Acquisition, and Operation
Abstract
(Approved for Public Release)Research Problem: The main goal of the proposed transformative research effort is to develop, atthe fundamental level, a disruptive Computational Science for building a new generation of numericalmodels that can accompany a DoD platform from its design, to its acquisition # which includes test& evaluation, as well as maintenance # to its operation; and can enable its predictive rather thanpreventive maintenance. The research effort will target three different levels of numerical modeling:a lower level pertaining to the material of which a platform is made, or to the environment in which itoperates; a middle level relating,among others, to the design, analysis, optimization, and acquisitionof the platform; and a higher level targeting its operation, optimal control, and predictive maintenance.Technical Approach: At the lower level, the technical approach will be grounded in the enablementof mechanics- and physics-informed deep artificial neural networks (ANNs) for constitutive modeling.At the middle level, it will be anchored in the construction of nonlinear approximation manifoldsbased on the scalable composition of affine subspace approximations rooted in data compressionalgorithms and nonlinear approximations of closure errors grounded in deep ANNs, kernel regressors,or finite-dimensional feature maps, to shatter the Kolmogorov barrier to parametric, projection-basedmodel order reduction (PMOR); in a radically reimagined PMOR that embraces models and data ofdifferent modalities, and transfer learning; and in innovative methodologies for breaking the curse ofdimensionality associated with training PMOR and surrogate modeling in general in high-dimensionalparameter domains. At the higher level, the technical approach will leverage data fusion to developgroundbreaking physics-based machine learning (PBML) procedures for extracting from test data,operational data, and/or high-dimensional numerical data knowledge or information not captured bya deterministic numerical model, and infusing it into a stochastic counterpart, to model and quantifymodel-form uncertainty, perform model updating, and enable transfer learning.Outcomes: The anticipated outcomes of the proposed transformative research effort will include newparadigms for data-driven constitutive modeling that feature the form-agnostic advantage of purelyphenomenological regression models, and the physical soundness of mechanistic and thermodynamicmodels. They will also include disruptive PMOR and other types of PBML methods that could supersedecurrent contenders for the real-time solution of a wide variety of steady (static) and unsteady(dynamic) problems in Computational Physics; can embrace multi-physics models and problem formulationswhose settings are dictated by requirements rather than limitations; can handle data ofdifferent modalities; and can cope with transfer learning to be able to accompany, for example, abuilding block approach. The anticipated outcomes will also include seminal probabilistic learningmethods that enable numerical models to continually evolve and update themselves in order to bettermatch reality andimprove the likelihood of achieving challenging objectives.Impact: Collectively, the anticipated outcomes highlighted above will revolutionize the modelingand simulation of a DoD platform, its design, acquisition, optimal control, and preventive ratherthan scheduled maintenance; will enable avant-garde PBML that overcomes effectively data scarcity;and will bridge the gaps between design, analysis, test & evaluation, operation, and the discoveryof operational anomalies. They will also enable the PI and his research team tomake significantcontributions to the education and training of the workforce in Computational Science, both in theclassroom and by reaching out and collaborating with researchers at DoD laboratories.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 13, 2023
- Source ID
- N000142312877
Entities
People
- Charbel Farhat
Organizations
- Office of Naval Research
- Stanford University
- United States Navy