Basic reporting

No comment. Experimental design

No comment. Validity of the findings

The work does not support the conclusion that this model of sprinting is suitable for research on sprint performance. At the very least, the limitations should be discussed. See general comments for details. Comments for the author

This manuscript presents simulations of sprint running, developed using a model of multibody dynamics with contact. The joint torques and contact model parameters were optimized to track kinematic data, measured ground reaction forces, and measured joint torques. It was found that the simulations could closely reproduce the data, with low dynamic inconsistencies. The authors concluded that the model is ready to be used for predictive simulations to study the cause-effect relationships in sprint performance. The paper was technically sound, well written and a pleasure to read.

I agree with the authors that a data-tracking optimization, as presented here, is a necessary and important test to determine whether a model has the capability to replicate a measured performance. Until a model passes this test, applications and predictive simulations should not be considered. However, I feel that this capability was not sufficiently tested, see comments 1 and 2, below.

1. The model is actuated by joint torques rather than muscles. Muscle-actuated simulation is already state of the art, for example, in the work by Falisse which is cited in the manuscript, and which shares one co-author with this manuscript. Muscle actuation would constrain the joint torques to the dynamic force-generating capabilities of muscles. Without constraints on the torques, almost any movement is possible, making the findings hardly surprising. This also is important for the intended application. If one wanted to study sprint performance, performance is only limited by arbitrary bounds placed on the joint torques, rather than bounds informed by muscle physiology. It is unlikely that such a model would provide insight into the factors that affect sprint performance. This limitation diminishes my enthusiasm for this work, and should, at very least, be properly discussed.

2. In the Introduction, the authors mention correctly that dynamic residuals should not be neglected in sports biomechanics research. However, their model still allows non-zero residuals, although they are minimized during the optimization process. If residuals are allowed, the movement is no longer strictly constrained by the laws of physics. Trajectory optimization with zero residuals is the state of the art, (e.g. the cited papers by Falisse) so there is hardly a good excuse for allowing residuals. This is also important for the envisioned applications. If a predictive simulation allows an external “helping hand” force acting on the pelvis, this becomes a major confounding factor for performance (running speed). The residuals that were found to be acceptable in this paper (34 Nm RMS in anterior-posterior direction, Table 3) are as large as the air drag force during sprinting, which is known to be a major factor in performance. I suspect that removing the external pelvis actuation would still produce very good results, and I strongly recommend making that change.

3. The description of the model did not mention air drag, so I assume it was not included. Air drag produces a significant amount of force during sprinting. A simple lumped force, proportional to squared velocity, could be modeled, for instance, based on Quinn, Journal of Sports Science, 2004. The lack of air drag in the model may well be the reason for the large anterior-posterior force residuals (Table 3). This is also important for the intended application: without air drag, performance will be overestimated. At the very least, this limitation should be properly discussed.

4. Line 201-202, “The generic modified model was linearly scaled to match the anthropometric and inertial characteristics of the athlete by using a measurement-based approach within OpenSim”. This requires a citation if it was a specific, previously described approach. If not, a more detailed description is needed.

5. Line 259. The optimization objective has five terms. The first three are data tracking terms, and it is logical to give them the same weight, with exception of the justified lower weight of the torque tracking term. The 4th term penalizes residuals, and it is not automatically clear that its weight (w1) should be the same as the angle and GRF tracking terms. I would prefer not to have this term at all (see comment 2), but if you have it, its weight should not be tied to the tracking term. I suggest defining w1 for angle and GRF tracking, w2 for torque tracking, w3 for residuals, and w4 for accelerations. Then define w = [0.1 0.01 0.1 0.0001]. This does not affect the results, but would prevent the reader from thinking that residuals need to be weighted the same as tracking data.

6. Line 273. If all variables were normalized to 10% of the range observed in the experiment, a constant factor 0.1 should be added in each denominator in the equation. This seems arbitrary. However, this would simply scale the whole objective function by a factor 100, and not affect the optimization problem in any way. Consider leaving out this mention of the 10% because it does not matter. If the scale factor of 100 was helpful to make IPOPT perform better, it can simply be mentioned as an overall scale factor, rather than applied to the denominator in each term. But be careful to account properly for the 5 mm normalization in the anterior-posterior translation variable.

7. Line 294-295. The optimization of the contact model parameters is a strong and innovative aspect of this work. The resulting contact model makes sense, and is probably going to improve the validity of predictive simulations. Consider highlighting this a bit more as one of the goals of the work.