<p>Spectral X-ray computed tomography enables material decomposition by leveraging energy-dependent X-ray attenuation properties. However, material decomposition with spectral CT requires a longer acquisition time to obtain sufficient numbers of photons in each energy bin. Sparse-view offers a practical solution to reduce acquisition time, but it introduces ill-posedness, degrading decomposition accuracy. This study introduces a material decomposition framework based on Neural Radiance Fields where material maps are represented using a multilayer perceptron (MLP). The material maps are then optimized through a spectral forward projection process based on the Lambert–Beer’s law, while a partition of unity (PoU) loss ensures the physical constraint on material maps. Our method was evaluated using simulated and real spectral CT datasets and compared with a traditional statistical approach. The results demonstrated that our method performs well in material decomposition under sparse-view conditions. The results suggest that our “neural material composition fields” framework offers accurate material decomposition robust to sparse-view conditions without requiring labeled training data.</p>