Accurate modeling of infrasound transmission loss is essential for evaluating the performance of the International Monitoring System network, which monitors compliance with the Comprehensive Nuclear Test-Ban Treaty. Among existing propagation modeling tools, the parabolic equation method enables transmission loss to be finely simulated using wind and temperature atmospheric models. However, its computational cost prohibits the exploration of a large parameter space in operational monitoring applications. To address this, a recent study exploited a deep learning algorithm capable of making ground-level loss predictions almost instantaneously, similar to parabolic equation-based outputs. However, this relied on interpolated atmospheric models, leading to an incomplete representation of the medium, and did not use the temperature as part of the inputs, which makes the network not adapted for long range propagation. In the current work, we address these limitations by using both wind and temperature fields as input, simulated up to 130 km altitude and 4,000 km distance. We also optimize several aspects of the network architecture. We exploit convolutional and recurrent layers to capture spatially and range-dependent features embedded in realistic atmospheric data, improving the overall performance. The network reaches an average error of 4 dB compared to full parabolic equation simulations and provides epistemic and data-related uncertainty estimates. Its evaluation on the 2022 Tonga-Hunga Ha'apai volcanic eruption demonstrates its generalization capabilities. The model adapts to a broad range of frequencies not included in the training, marking a significant step towards near real-time assessment of International Monitoring System detection thresholds of explosive sources.