This paper presents a novel ablation analysis methodology for deep learning-based physical side-channel analysis (SCA) to address the portability problem, which arises when profiling attacks on devices differ due to variations in hardware and measurement setups. The proposed approach introduces a layer assessment methodology based on the ablation paradigm to evaluate the sensitivity and resilience of each neural network layer. This methodology helps identify which layers are critical for the model's performance and which can be removed without significantly affecting the model's ability to generalize across devices. By ablating specific parts of the neural network and performing recovery training, the model can be adapted to different devices without requiring multiple copies of the training device. The paper introduces the Multiple Device Model from Single Device (MDMSD) approach, which allows a model trained on a single device to generalize to leakage traces from various devices. This is achieved by partially ablating a selected layer of the model and then conducting recovery training with perturbed leakages from the original device to simulate the portability effect. The resulting model can generalize to various devices without relying on multi-device assumptions. The methodology is evaluated using two neural network types (MLP and CNN) and four datasets. The results show that the MDMSD approach significantly improves the model's ability to generalize across devices, with performance comparable to the Multiple Device Model (MDM) approach. The paper also introduces two new layer assessment criteria: sensitivity and resilience. Sensitivity measures the importance of a layer for achieving high attack performance, while resilience represents the necessity of a layer for the overall neural network. The results demonstrate that even smaller neural networks can achieve top performance in SCA, indicating room for further improvement in network design methodologies. The proposed approach addresses the portability problem in SCA, where the training device and the device under attack differ. The MDMSD approach shows promising results in overcoming this challenge, making side-channel analysis more accessible and effective in practical scenarios. The paper concludes that the proposed methodology provides valuable insights into the functioning of neural networks and helps bridge the gap between single-device models and MDM approaches when multiple devices are unavailable.This paper presents a novel ablation analysis methodology for deep learning-based physical side-channel analysis (SCA) to address the portability problem, which arises when profiling attacks on devices differ due to variations in hardware and measurement setups. The proposed approach introduces a layer assessment methodology based on the ablation paradigm to evaluate the sensitivity and resilience of each neural network layer. This methodology helps identify which layers are critical for the model's performance and which can be removed without significantly affecting the model's ability to generalize across devices. By ablating specific parts of the neural network and performing recovery training, the model can be adapted to different devices without requiring multiple copies of the training device. The paper introduces the Multiple Device Model from Single Device (MDMSD) approach, which allows a model trained on a single device to generalize to leakage traces from various devices. This is achieved by partially ablating a selected layer of the model and then conducting recovery training with perturbed leakages from the original device to simulate the portability effect. The resulting model can generalize to various devices without relying on multi-device assumptions. The methodology is evaluated using two neural network types (MLP and CNN) and four datasets. The results show that the MDMSD approach significantly improves the model's ability to generalize across devices, with performance comparable to the Multiple Device Model (MDM) approach. The paper also introduces two new layer assessment criteria: sensitivity and resilience. Sensitivity measures the importance of a layer for achieving high attack performance, while resilience represents the necessity of a layer for the overall neural network. The results demonstrate that even smaller neural networks can achieve top performance in SCA, indicating room for further improvement in network design methodologies. The proposed approach addresses the portability problem in SCA, where the training device and the device under attack differ. The MDMSD approach shows promising results in overcoming this challenge, making side-channel analysis more accessible and effective in practical scenarios. The paper concludes that the proposed methodology provides valuable insights into the functioning of neural networks and helps bridge the gap between single-device models and MDM approaches when multiple devices are unavailable.