Can a chess engine use neural network?

 Yes. Modern chess engines routinely use neural networks, and the strongest ones rely on them heavily.

How they are used

• Evaluation function: Instead of (or in addition to) hand-crafted evaluation terms (material, pawn structure, king safety, etc.), a neural network scores positions. This is the most common and successful application.
• Move selection / policy: Some engines (especially those inspired by AlphaZero) use a network to suggest promising moves, guiding the search.
• Hybrid designs: Many top engines combine classical alpha-beta (or similar) search with a neural network evaluation. Pure neural-network approaches that rely mainly on Monte-Carlo Tree Search also exist.

Notable examples

• Stockfish (the dominant open-source engine) has used NNUE (Efficiently Updatable Neural Network) since 2020. The network is trained on millions of positions and is extremely fast to evaluate on CPU, which is why it fits well inside a traditional search framework.
• Leela Chess Zero (Lc0) and engines derived from the AlphaZero approach use deep residual neural networks trained by self-play reinforcement learning. They typically use GPU acceleration and MCTS-style search.
• Other engines (Komodo Dragon, certain commercial and experimental engines) have also incorporated neural-network components.

Neural networks do not replace search; they improve the quality of the evaluation and/or the move ordering that search relies on. The combination of deep learning evaluation + efficient search is what produces the extremely high playing strength seen in current top engines.