Towards reliable and fair probabilistic predictions: field-aware calibration with neural networks