The goals of this research were to 1) critically evaluate muscular models of, 2) critically evaluate neural models of, and 3) develop a novel method for generating results from, forward dynamic simulations. The objective of the muscular model evaluation was to determine whether the use of a hyperbolic, Hill-type muscle equation better replicates experimental data when the equation parameters (i.e. the ‘shape’ and the ‘V_max’ parameters) are dependent on the muscle activation, and the objective of the neural model evaluation was to determine whether the use of sloped excitation waveforms better replicate experimental data when compared to rectangular waveforms. These objectives were achieved in part by collecting EMG, kinetic, and kinematic data from subjects performing elbow extension and recumbent pedaling tasks.

Neuro-musculo-skeletal models were generated for both the elbow extension and recumbent pedaling tasks. The triceps brachii were modeled for the elbow extension task and nine functional muscle sets of the lower extremity were modeled for the recumbent pedaling task. Excitation waveforms were used as inputs to the forward models. These were transformed into activations which in turn were used as inputs to the muscle models to simulate muscle forces. Due to dynamic coupling both the body segments and joints were accelerated by these forces, producing a simulated motion that included the joint kinematics and kinetics.

To evaluate the muscular model subjects performed rapid-release elbow-extension tasks at different activation levels. The errors between experimental and simulated angles were minimized by optimizing the Hill parameters (V_max and k) in the model, and these values were regressed onto the activation to assess for dependencies on the activation level. Although the shape parameter did not demonstrate a dependency on the activation level, the V_max parameter did, with V_max increasing with activation levels.

To evaluate the neural models subjects performed both elbow extension and recumbent pedaling tasks. Four models were evaluated, one implementing a rectangular excitation waveform and three implementing sloped waveforms (triangular, quadratic, and sinusoidal). The errors between experimental and simulated kinematics and kinetics were minimized by optimizing the excitation waveform parameters. The relative accuracies of the waveforms were assessed by comparing tracking errors and muscle timing errors. Models using sloped waveforms generally produced both smaller tracking and smaller timing errors compared to rectangular waveforms, even though the model complexity was constant across waveforms (each waveform was defined by three parameters).

The main impetus for developing the novel method for generating simulation results was to be able to statistically validate the simulation results. Specifically, the objectives were to:​ 1) generate a distribution of simulation results, the means and deviations of which could be used to statistically validate the simulation results, 2) perform such a validation, and 3) explore any model inaccuracies manifested with the method. These objectives were achieved in part by collecting EMG, kinetic, and kinematic data from subjects performing the recumbent pedaling tasks, and from implementing a neuro-musculo-skeletal model of recumbent pedaling.

The method continually generated new simulation results that were included in a distribution of simulation results. A vector-valued objective function, defining the errors between simulated and experimental data, was implemented to generate a Jacobian which was used to define control parameter directions that tended to decrease the objective function components. The control parameter directions and step sizes were periodically updated so that the objective function components were continuously increasing and decreasing between simulations. This process of computing the control parameter directions and step sizes, performing simulations to include in the distribution, and then re-computing the control parameter directions was continued until the distributions of muscle quantities of interest (average muscle activation, force, and work) stabilized.

The simulated deviation patterns (e.g., magnitude over the crank cycle) were in general agreement with experimental deviation patterns;​ however the simulated means differed considerably from the experimental means. The errors in the means, explicitly attenuated in the solution from an optimization method, were traced to inaccurate components in the neuro-musculo-skeletal model. One type of error in the means was traced to a common simplification of the neuro-musculo-skeletal model (assuming left-right symmetry). Using the distribution of simulation results, high frequencies in the simulation outputs, not present in the experimental output, were associated with muscle velocities near zero. These results suggested that muscle history should be considered in forward simulation models to increase the accuracy of the simulation results.

This results of this work may be used to:​ 1) increase the accuracy of forward simulation analyses (by implementing refined neural and muscular models), 2) generate statistical confidence of forward simulation results (by implementing the distributiongenerating method), and 3) investigate sources of inaccuracies in neuro-musculo-skeletal models (by analyzing the results from the distribution-generating method).