Machine learning tools and techniques have been dramatically improved over the last years and have a large potential of use in digital soil mapping. The aims of this study were (i) to predict soil parent material, an important pedogenetical factor, and (ii) to use a machine learning method using punctual soil descriptions as training data. We used MART (Multiple additive and regression trees; Friedman, 1999) as machine learning system. It uses stochastic gradient boosting, a particular boosting method known to improve the accuracy of regression and classification trees. Contrary to previous studies, we tested the use of punctual observations as training data instead of soil maps, which enables a more even distribution of the training data over the region of interest. The study area was Brittany (NW France), a region of 27 020 900 ha. In this region, bedrock comprises mainly metamorphic and eruptive substrates. However, superficial deposits, namely Aeolian loam, represent very often the actual soil parent material and are very poorly delimitated by existing geological or geomorphological maps. MART requires two kinds of input data to calibrate a predictive model: training data, which correspond to the response variable to predict, and environmental predictors. 5 129 profiles (punctual observations of soil), were used as training data. Response variable were composed of 11 types of geological bedrocks and 9 types of superficial deposits. We used 17 environmental predictors (terrain attributes derived from a 50 m resolution DEM, emissions of potassium obtain by airborne gamma-ray spectrometry…). Several validations were performed. Model predictions were compared to (i) punctual data, used or not to calibrate the model (respectively internal and cross-validation), and (ii) to existing 1/25 000 soil maps (external validation).Internal and cross-validation of the model showed an accuracy of parent material prediction 81% and 54%, respectively. External validation accuracy was 44%. The most important predictors were bedrock lithology, landscape map, change in altitude and Beven topographic index. Best predicted materials were granite, brioverian shale, schist, Aeolian loam and sand dune. Uncommon materials and colluvium were found to have the highest error rates. Poor prediction was also associated with areas with high geological complexity. Results showed that machine learning from punctual soil observations has a high potential and should be more studied. The resulting predictive map of soil parent material will be a basis for prediction of further soil properties, namely soil waterlogging.