Full title: Karan Taneja - Feature Encoded and Multi-Resolution Physics-Informed Machine Learning Approaches for Musculoskeletal Digital Twin Applications

Machine Learning (ML) approaches offer effective tools to identify biological system properties from physiological measurements, e.g., motion data and raw surface electromyography (sEMG), providing opportunities to construct a subject-specific musculoskeletal (MSK) digital twin system for health condition assessment and motion prediction. While physics-informed ML approaches for dynamic systems offer learning capabilities that satisfy the conservation laws, physics-informed time-domain mapping of high-frequency muscle excitation signals to low-frequency joint motion remains challenging owing to the large variation in frequency contents between the activation signals (input) and motion data (output). In this work, we first developed a Feature-Encoded Physics Informed Parameter Identification Neural Network (FEPI-PINN) for the simultaneous prediction of motion and parameter identification of human MSK systems. Here, the features of high-dimensional noisy sEMG signals were projected onto a low-dimensional noise-filtered embedding space for effective forward dynamic training. This FEPI-PINN model can be trained to relate sEMG signals to joint motion and simultaneously identify the key MSK parameters. To enhance time-domain mapping, we then propose a Multi-Resolution Recurrent Neural Network (MR-RNN) learning algorithm. In this approach, the fast wavelet transform is applied to noisy sEMG signals, decomposing them into nested multi-scale signals. The prediction model is first trained with lower-resolution input signals using a gated recurrent unit (GRU), and the trained parameters are then transferred to the next higher-scale training. These training processes are repeated recursively until a full-scale training is achieved. Numerical examples demonstrate that the proposed framework can effectively identify subject-specific muscle parameters with noisy sEMG signals, and the trained physics-informed forward-dynamics surrogate yields accurate motion predictions of elbow flexion extension motion, which are in good agreement with the measured joint motion data.