Bayesian-Inference-TRPL

A Bayesian inference algorithm and script for the fitting of TRPL observations from absorbers to the carrier drift/diffusion/decay model described in Minority Carriers in III-V Semiconductors: Physics and Applications (Ahrenkiel and Lundstrom, 1993) and Chalcogenide Photovoltaics: Physics, Technologies, and Thin Film Devices (Scheer and Schock, 2011).

File Overview

• parallel_bayes_gpu.py - Entry script and configuration settings for the inference. Do python parallel_bayes_gpu.py OBSERVATION_FILE EXCITATION_FILE OUTPUT_NAME [new|new+|load] to start the show.
• bayeslib.py - Driver functions for generating random samples of the parameter space, delegating simulations to GPUs, and computing resultant likelihoods.
• pvSimPCR.py - GPU solver for TRPL simulation. Uses parallel cyclic reduction. Can be run directly as a standalone simulator when provided with an input pickle.
• probs.py - GPU-accelerated likelihood computations.
• bayes_io.py - Utility functions for loading OBSERVATION and EXCITATION files and exporting output likelihood files.
• bayes_validate.py - Validation for configuration settings in parallel_bayes_gpu.py.
• Example Data
  • Power_scan_Excitations.csv - Excitations/Initial Conditions for the "exemplary" three-level power scan shown in the article.
  • Power_scan_Observations.csv - Observations/TRPL data points for the "exemplary" three-level power scan shown in the article. Goes with Power_scan_Excitations.csv.
  • Highsurf_Power_scan_Observations - Observations/TRPL data points for the "alternate high surface" power scan shown in the article. Also uses Power_scan_Excitations.csv.
  • Twothick_Excitations.csv - Excitations/Initial Conditions for the two-thickness scan discussed in the article.
  • Twothick_Observations.csv - Observations/TRPL data points for the two-thickness scan discussed in the article. Goes with Twothick_Excitations.csv
• Testing
  • pvSetup.py - Generate an input pickle for running pvSimPCR.py standalone.
  • PV_tester2.py - Simulate TRPL using input pickle and scipy PDE solvers. Generates output pickle containing simulation results.
  • compare.py - Compare two output pickles. Perhaps one from pvSimPCR.py and one from PV_tester2.py?
  • pvPlt_interface.py - Basic visualizer for output pickles.
• Legacy
  • legacy.py - Functions for coarse grid sampling of the parameter space.
  • parallel_bayes.py - Serial CPU version of the Bayesian inference algorithm.
  • pvSim.py - Serial CPU version of the TRPL simulation solver.

Usage

1. Configure parallel_bayes_gpu.py. Adjust simPar, parameter space boundaries, ic_flags, gpu_info, sim_flags, init_dir, and out_dir.
2. Prepare OBSERVATION and EXCITATION files. The format of these are as follows: OBSERVATION files are three column lists of TRPL data points: timestamp, TRPL value, and experimental uncertainty value. If multiple TRPL curves are to be inferenced, the sets of data points for each curve should be concatenated one below the next. Add a single line with the word "END" to the end of the list. EXCITATION files are lists of initial carrier density profiles - each row corresponding to one curve in the OBSERVATION file. See Example Data for exmaples.
3. Run python parallel_bayes_gpu.py OBSERVATION_FILE EXCITATION_FILE OUTPUT_NAME [new|new+|load].
4. Collect "BAYRAN_X" parameter sample file and "BAYRAN_P" raw likelihood file.
5. Compute posterior probability distribution from "BAYRAN_X" and "BAYRAN_P". Example script coming soon!