The manuscript entitled "Structural origin of disorder-induced ion conduction in NaFePO4
cathode materials" attempts to reveal structural origin of maricite cathode material amorphization through MD simulations enabled by machine-learning potentials. The authors are suggested to address the following comments:

1. Transferability and Structural Optimization Criteria:
   The authors state that the developed ML potential is exposed to a wide range of chemical environments to enhance its transferability and stability. However, it appears that structural optimization is performed primarily with respect to energy (and possibly forces), which may not adequately capture other critical material properties. It is well-documented that many ML potentials trained on large datasets can achieve low errors in energy and force predictions, yet still fail to accurately reproduce higher-order properties such as the elastic tensor. Including experiments on additional compositions would help strengthen the claims regarding transferability and robustness.

2. Use of Density for Structural Validation:
   The authors use density as a primary metric for validating the optimized structures. However, since different compositions can yield similar densities, this metric alone may not provide a sufficiently discriminative or robust basis for validation. Would the authors consider including additional physical properties—such as elastic moduli or other mechanical parameters—to more comprehensively assess the validity of the optimized structures? Incorporating such properties could offer stronger evidence for the accuracy and physical relevance of the ML potential beyond matching density alone.

3. Choice of ACE and Architectural Considerations:
   The authors employ the ACE framework, which is understandable given its integration within the PACEMAKER environment. However, ACE is not a fully equivariant model, and training invariant models on limited datasets often poses generalization challenges. In this context, could the authors comment on the adequacy of the dataset size and chemical diversity? Additionally, have the authors considered comparing their results against fully equivariant architectures such as MACE, MatterSim, or SevenNet, which have recently demonstrated improved performance in capturing complex physical symmetries and ensuring stability?

4. Optimization Strategy and Model Validation:
   The use of the BFGS algorithm for structure optimization is acknowledged. While BFGS is a well-established gradient-based method, Bayesian optimization has shown superior performance in high-dimensional search spaces. Would the authors consider benchmarking the performance of BFGS against a Bayesian optimization approach to evaluate potential improvements in structure prediction or model robustness?