Alan Salcedo

MIT Department: Physics

Undergraduate Institution: University of Texas, El Paso

Faculty Mentor: John Negele

Research Supervisor: Phiala Shanahan, Andrew Pochinsky

Websites: LinkedIn, Personal site

Biography

I am a Physics major aiming to obtain a Ph.D in Nuclear Physics. I was born and raised in Ciudad Juarez, Chihuahua, Mexico. In order to pursue my education, I cross the Ciudad Juarez – El Paso boarder and attend classes at the University of Texas at El Paso (UTEP). My ultimate goal is to perform research at a world-wide renowned facility, such as Los Alamos National Laboratory or Oak Ridge National Laboratory, as well as to teach multiple Physics courses at a university. I enjoy science and working in community projects.

2017 Poster Presentation

2017 Research Abstract

Classification of Dynamical Systems and Prediction of their Physical States Using Deep Learning

Alan Salcedo, Department of Physics, University of Texas at El Paso
Phiala Shanahan, Center for Theoretical Physics, Massachusetts Institute of Technology
Andrew Pochinsky, Center for Theoretical Physics, Massachusetts Institute of Technology
John Negele, Center for Theoretical Physics, Massachusetts Institute of Technology

Deep learning is associated with predicting the state of a physical system without directly solving its equations of motion. Recent studies have made it possible to produce computer programs able to create algorithms for predictions using large datasets. These programs are designed based on networks of neurons and are called Deep Neural Networks (DNNs). These programs are often efficient for calculations that are time consuming, such as those in Lattice Quantum Chromodynamics (LQCD). In this study, we tested deep learning methods in the classification of classical oscillators and in the prediction of the evolution of their coordinates and momenta. We created a DNN with 4 fully connected hidden layers of 32 nodes capable of classifying these systems and predicting their trajectories for any parameters and initial states. This network was trained with artificial classifier arrays and sets of 300,000 trajectories of the Simple Harmonic Oscillator (SHO), Simple Pendulum (SP), and Double Pendulum
(DP) produced by numerical solutions to their equations of motion. We calculated the
difference of its predictions to the actual solutions for 300 test points on each task. For time evolutions of 0.01, 0.1, 1, and 3.2 sec, the DNN was able to differentiate the SP and SHO up to 90% on average with standard deviations of 0.2. For time evolution of 0.01 sec, the predictions of the trajectories of the SHO departed 0.039 units from the actual solutions with a standard deviation of 0.025. Corresponding results were obtained for every configuration. We intend to further improve the DNN to reduce
the error of its predictions and investigate its efficiency in calculations of LQCD.