Machine Learning modeling of the stopping power experimental data

C.C. MONTANARI; F. BIVORT HAIEK; MENDEZ, A.M.P.; D. M. MITNIK

Congreso; INTERNATIONAL CONFERENCE ON ION BEAM MODIFICATION OF MATERIALS; 2022

Escola Superior de Tecnologias da Saúde de Lisboa

The International Atomic Energy Agency (IAEA) stopping power database [1] is a valuable publicresource, which is continuously updated [2], making available compilations of the experimentalmeasurements published over the last nine decades [3] to the global scientific community. The purposeof this work is to apply machine learning methods to predict the electronic stopping power cross sectionbased on this important compilation of data. The experimental data compiled from published articlesby diverse authors is not standardized. The data is presented in distinct units and formats, including themix of stopping cross sections per atom, per molecule, per mass and also energy loss per unit pathlength. We devoted a significant effort in the reorganization of the database, unifying units, andarranging the data in a standard (csv) format, enabling to obtain the information easily and quickly.Considering that the database contains several dozen of thousands input values, purging manually thisdata requires also a considerable amount of work.To this purpose, we developed a machine learning method to clean up the existing compilation ofexperimental data, based on an unsupervised clustering technique (DBSCAN) which identifies outliervalues, and determines which data to keep in cases of overlapping, taking into account the year ofproduction. The refined data is used to train a deep neural network, able to accurately reproduce theinput experimental results, and to predict new results in cases which have never been seen in the trainingprocedure (the test set). In this opportunity, the code ESPNN (Electronic Stopping Power with NeuralNetwork), performing all these tasks will be presented, and the results discussed.