High-spectral-resolution atmospheric modelling and physical retrieval need fast, differentiable infrared radiative transfer to support repeated forward simulations and accurate Jacobians at large scale. Line-by-line models are the accuracy reference but prohibitively slow for practical applications. We introduce this PCOD-NN (principal-component optical-depth neural network), a neural-network-based emulator that reproduces the layer-by-layer, gas-by-gas monochromatic optical depth (OD) over 600–2600 cm^−1 at 0.01 cm^−1 resolution for seven gases (H₂O, CO₂, O₃, N₂O, CO, CH₄, and O₂). Instead of fitting radiance of a particular instrument, PCOD-NN predicts OD itself—a quantity upstream of radiative transfer (RT) and instrument convolution—so the same representation couples to different RT solvers, instrument line shapes, and spectral response functions. For each gas, ln(OD) is compressed onto a 500-component PCA basis whose scores are predicted by a residual multilayer perceptron, with separate high- and low-pressure models over 1–1100 hPa. On held-out test sets, the mean relative OD error is less than 0.2 % for each of the seven gases. For 90-layer profiles, the V100 GPU implementation is approximately 620–970 times as fast as single-core LBLRTM in the full-resolution seven-gas OD benchmark. The predicted OD spectra are then used in a clear-sky radiative-transfer and instrument-convolution pipeline executed on the GPU to calculate channel brightness temperatures. For the FY-4B/GIIRS hyperspectral infrared sounder, agreement between simulated and observed brightness temperatures is overall comparable to that achieved by RTTOV v14. Without retraining the OD emulator, the framework also simulates infrared channels of the FY-4B/AGRI multispectral imager in a clear-sky ocean scene, achieving agreement with observations broadly comparable to that achieved by RTTOV v14 and demonstrating cross-instrument reuse. Across 12 diverse atmospheric profiles, temperature and water-vapor brightness-temperature Jacobians obtained by automatic differentiation through the forward model have median relative errors of approximately 0.2 % and 0.3 %, respectively, relative to finite-difference references computed using LBLRTM. By combining accurate emulation of high-resolution, gas-resolved optical depth with differentiable radiative transfer, PCOD-NN provides a common framework for hyperspectral sounding, broadband imaging, and atmospheric sensitivity analysis.