Implementation of Deep Learning in New Particle Searching

Project Check in #1 (Outline)

Title: Implementation of Deep Learning in New Particle Searching.

Who: Chongwen Lu, Yidong Zhou.

Introduction: In some Dark Matter Searching Experiments (eg. LUX-Zeplin), when particles enter the detector, they will collide with the liquid inside the detector, the collision events include multiple scattering and single scattering, etc. Single scattering (SS) signal has a shape of Gaussian, multiple scattering (MS) signal consists of two or more Gaussians. Some MS signals look very similar to SS signals, we will need a fast and accurate way to distinguish them. This can help to improve Dark Matter detection sensitivity. We will use distinguishing SS and MS is a classification problem, but fitting Gaussians inside a SS or MS is a regression problem. We will use supervised learning to train our model.

Related Work: We found three papers that might be useful to our project, we are looking at more. Fast curve fitting using neural networks: https://aip.scitation.org/doi/pdf/10.1063/1.1143696 Integration of Multi-Gaussian fitting and LSTM neural networks for health monitoring of an automotive suspension component: https://www.sciencedirect.com/science/article/abs/pii/S0022460X18302815?via%3Dihub Multi-Gaussian fitting for pulse waveform using Weighted Least Squares and multi-criteria decision making method: https://pubmed.ncbi.nlm.nih.gov/24209911/

The first paper (Fast curve...) introduces a way of fitting multiple Gaussians using MLP. This method is much faster than iterative methods and does not require an initial guess for the solution. The error histogram generated from fitting result shows that the majority of fittings on 1000 simulated signals have errors less than 3%.

Data: We will use simulated signal that is superposed of Gaussian signals based on detector settings, this can perfectly mimic the Poisson noise of real LZ signal. Each signal will consists of around 500 bins, we will need approximately 10000 signals for training, .etc. As we create the dataset ourself, no significant preprocessing will be needed.

Methodology: We will use simulated data to train my model so that we know the actual parameters of the Gaussians in signals. This will be a supervised learning process. Our initial thought is that we will first implement and test the method mentioned in "Fast curve fitting using neural networks" to make initial guesses, then we will use LSTM-FCN to make classification and fitting, this should lead our fitting to higher accuracy. We will try LSTM because the signals in LZ experiment will have certain patterns. LSTM, which is good at making use of "context" should be a good candidate.

Metrics: I will try to run two experiments to test the successfulness. The first experiment is to test the accuracy of classifying MS and SS, I hope we will get an error rate of less than 100 PPM (part per million). The second experiment I will try to run is comparing the sigma values from fitting results and actual sigma values, I will hope the majority (98%) of tests have an error of less than 5%.

Ethics: Why is Deep Learning a good approach to this problem? Traditionally we use least square method to fit signals with one or multiple Gaussians, the results are sensitive to initial guesses. While noise level is high, it is hard to make correct initial guesses. Also least square method is very slow compare to (some/proper) deep learning methods. What is your dataset? Are there any concerns about how it was collected, or labeled? Is it representative? What kind of underlying historical or societal biases might it contain? The dataset is created by a simulation code written by Chongwen. There is no concerns about how it was collected. However, due to some restrictions, we are not able to use real LZ data in this project. Also we are not able to publish the simulation code as it contains some detector settings.

Division of labor: Chongwen will create simulated dataset. Chongwen and Yidong will design and train the model together.