This thesis investigates neural network-based approaches for the inverse calibration of stochastic volatility models, with a focus on both predictive performance and structural interpretability. Calibration is formulated as an inverse problem, where model parameters are inferred from implied volatility surfaces, a setting that is inherently non-linear and potentially ill-conditioned. While neural networks offer a computationally efficient alternative to classical optimisation-based methods, existing work has primarily focused on predictive accuracy, with comparatively limited attention given to the interpretability of the learned inverse mappings. This lack of transparency raises concerns in financial applications, where understanding model behaviour is essential for validation and risk management. Recent work has begun to address this issue, but a systematic framework for validating the learned inverse mappings remains limited. To address this, we evaluate several neural network architectures, including the multilayer perceptron (MLP), Highway networks, and Generalised Highway networks, trained on synthetically generated implied volatility surfaces from the Heston and rough Heston models. We introduce a framework for the structural validation of learned calibration mappings using eXplainable Artificial Intelligence (XAI) methods, specifically SHAP and νSHAP, which capture complementary notions of feature relevance.

The results show that neural networks achieve highly accurate and fast calibration within the training domain, with the Generalised Highway architecture consistently outperforming its counterparts in terms of validation error. The XAI analysis confirms that the learned mappings rely on structurally meaningful regions of the implied volatility surface, most notably short maturities and the wings of the smile, aligning with financial intuition. At the same time, the comparison between SHAP and νSHAP highlights non-homogeneous parameter identifiability: while some parameters are associated with concentrated and stable regions of importance, others are inferred from diffuse and overlapping information, indicating redundancy in the inverse mapping.

In addition, we demonstrate that XAI can be used to inform model design. In particular, a νSHAP-guided reduction of the input grid yields comparable or improved predictive performance despite a substantial reduction in input dimensionality.

Overall, the findings show that neural network-based calibration can be both accurate and structurally interpretable when combined with appropriate analysis tools, while also revealing important limitations related to generalisation and parameter identifiability.