A tunable mid-infrared (MIR) laser (quantum cascade laser, QCL) was used for the detection of TNT and RDX in soil samples at a concentration range from 0 to ~20% w/w. This type of sensing is complicated due to the complexity of the matrix, i.e., the diversity of compounds contained in soil. Thus, the high explosives (HE) detection in soil by QCL was assisted with an Artificial Intelligence (AI) system. AI managed to predict these HE in seven kinds of soils using minimum information Machine Learning (ML). The models were generated only from neat HE and soil spectra, without necessity using experimental spectra of the mixes. AI used these neat spectra to simulate the spectra of HEs/soil mixes. The simulated data was used to train the ML models and then were tested with real spectra of HEs/soils mixes. The method was designated as “Self-Simulated Learning Artificial Intelligence” (SSLAI). This methodology has advantages for applications in field scenarios where the matrices are unknown because SSLAI models do not need to be trained with real samples a priori. Models would only have to be fed with spectra for the neat components to train itself. The methodology was tested with mixes of seven soils and two explosives. Test samples were classified into three concentrations ranges: high concentration test (Test_H > 10% w/w), medium concentration test (10% w/w > Test_M > 3% w/w), and low concentration test (Test_L < 3% w/w). The results show that it is possible to correctly predict these two HE/soil mixes from the simulated data. Specifically, for TNT and RDX, SSLAI achieved a high precision in the prediction for the high and medium concentration tests (Test_H and Test_M). However, for both samples with concentrations below 3% w/w (Test_L), the number of false positives increased, and the precision was reduced.