This project aims to establish declarative combinatorial optimisation as a framework for training neural networks with explicit, user-defined requirements. While neural networks have achieved impressive predictive performance, conventional training typically expresses learning goals through numerical loss functions. Incorporating logical rules, structural restrictions, and competing requirements can therefore be difficult. Building on the substantial success of declarative methods in solving complex combinatorial optimisation problems, this project will investigate how their modelling expressiveness and search capabilities can contribute to neural network learning. The central idea is to describe the learning task in terms of what a network should achieve and which conditions it should satisfy, allowing optimisation methods to search for suitable solutions. This creates opportunities to integrate domain knowledge directly into training and to explore requirements that are difficult to capture through a loss function alone. The research will consider both the development of new learning formulations and the adaptation of existing optimisation techniques. The project will investigate constraint programming, satisfiability-based methods, mixed-integer optimisation, and combinations of these approaches. Hybrid methods that integrate declarative reasoning with numerical optimisation will also be explored. A central research question is how to balance expressive modelling, predictive performance, and computational tractability as the size and complexity of the learning problem increase. Where existing methods are insufficient, the project may develop specialised algorithms or solver components. Robustness, fairness, and explainability will serve as important research directions within this broader framework. Potential applications include enforcing specified behavioural properties, limiting disparities according to selected fairness criteria, and encouraging interpretable network structures. Explicit constraints will make requirements and modelling assumptions inspectable, while explanations of predictions and constraint violations will require additional investigation. Any formal guarantees will be restricted to the encoded properties and their stated assumptions. The methods will be evaluated through systematic comparisons across benchmark tasks, network architectures, optimisation formulations, and solvers. Evaluation will examine predictive accuracy, constraint satisfaction, relevant robustness and fairness measures, and computational efficiency. Particular attention will be given to conflicting requirements and to identifying when declarative approaches offer practical advantages over established training methods. The expected outcomes are new formulations, algorithms, and a reproducible experimental framework for training neural networks through declarative combinatorial optimisation. By connecting advances in combinatorial optimisation with machine learning, the project seeks to make learning systems more controllable and adaptable to explicit human requirements, while establishing the capabilities and limitations of this emerging research direction.