To address the low training efficiency of the VQE algorithm, this paper proposes Titan, a deep learning-based framework. Titan identifies and fixes redundant parameters during initialization for a given Hamiltonian, thereby reducing optimization overhead while maintaining accuracy. This is based on the empirical finding that some parameters have minimal impact on training dynamics. It is designed by combining an informative and Barren Plateau-resistant training data generation strategy with an adaptive neural network architecture that generalizes to Ansatz of various sizes. Benchmarks on Ising models, Heisenberg models, and various molecular systems with up to 30 qubits demonstrate that Titan achieves up to 3x faster convergence and 40-60% fewer circuit evaluations than existing state-of-the-art baseline models, while achieving comparable or superior accuracy.
