Data science and Quantum Information Science division focuses on sharing knowledge, state-of-art tools, techniques, research, and emerging industry trends with machine learning applications. Machine learning tools are becoming prevalent with a growing impact across industries. Reflecting the increased relevance, this division aims to create a forum to exchange experiences and discussions on research topics and inspire collaborations bringing people together from academia to industries.
Physicist’s Journey to Machine Learning Fraud Prevention Scientist
The transition from a career in physics to one focused on machine learning for fraud prevention in financial systems represents a compelling narrative of interdisciplinary exploration and professional evolution. This abstract offers insights into this journey, highlighting the unique perspectives, challenges, and opportunities that physicists encounter when venturing into the realm of solving the real-world problem of fraud prevention through the machine learning model.
We physicists possess a diverse skill set encompassing advanced mathematical techniques, computational modeling, and a deep understanding of complex systems dynamics. As they pivot towards addressing the intricate challenges of fraud detection and prevention, physicists leverage their analytical prowess to uncover patterns, anomalies, and underlying structures within vast datasets of financial transactions.
This also includes the methodologies that physicists can employ for transitioning to machine learning scientists, emphasizing integrating physics principles into algorithmic design and model development. Drawing parallels between concepts in statistical mechanics, network theory, and quantum mechanics, physicists can craft innovative approaches to identify fraudulent activities with heightened accuracy and efficiency.
Furthermore, this also explores the pivotal role of interdisciplinary collaboration in facilitating this career transition. Physicists engaging with data science, cybersecurity, and finance experts should gain domain-specific knowledge and refine their understanding of fraud dynamics in financial ecosystems. Through collaborative efforts, physicists can enhance their ability to translate theoretical insights into practical, real-world solutions that mitigate risks and safeguard financial institutions and consumers.
This also reflects on the personal and professional growth experienced by physicists navigating this transition. From acquiring new technical skills and domain expertise to embracing challenges and cultivating adaptability, the journey from physicist to machine learning fraud prevention scientist fosters a spirit of continual learning and innovation.
In conclusion, this underscores the significance of interdisciplinary collaboration and the potential for cross-pollinating ideas between physics and machine learning in addressing real-world challenges such as fraud prevention. By embarking on this career transition, physicists can contribute their unique perspectives and analytical rigor to advancing the state-of-the-art in fraud detection and prevention, ultimately strengthening the resilience of financial systems in an increasingly digital age.
Please look below for detailed schedule.
|
Date/Time: |
Abstract Number: ANPA2024-N00033 Presenting Author: Keshab Chaudhary, Dipendra Prasad Kalauni Presenter's Affiliation: Central Department of Physics , Tribhuvan University Title: Machine Learning Driven Formation Energy and Lattice Parameter Prediction of Full Heusler Alloys: X2YZ Location: Poster Presentation Show/Hide Abstract
Machine learning, a part of both AI and data science, aims to enable machines to recognize patterns and generate decisions or predictions without requiring explicit programming. This work is based on the application of supervised learning in material science, utilizing machine learning as a useful tool for quickly identifying new materials and predicting their features. To predict the essential formation energy and lattice parameter, two supervised learning regression approaches within machine learning—random forest regression and gaussian process regression—have been used. In this study, the formation energy and lattice parameter of 140 full Heusler compounds were predicted by selecting 24 features of cubic structures with an 80% and 20% train-test split. Features such as volume, atomic radii, atomic mass, electro-negativity, lattice angle, band gap, formation energy, covalent radii, density, total magnetization, number of atoms, etc. were employed to predict the lattice parameter of cubic Heusler compounds (X2YZ). Model performance was evaluated using evaluation criteria such as mean squared error (MSE), mean absolute error (MAE), and R-squared (R2). After using both models, the obtained values for formation energy were as follows: Random Forest Regression—R2: 0.658, MSE: 0.011, MAE: 0.091, Gaussian Process Regression—R-squared: 0.734, MSE: 0.246, MAE: 0.396, indicating the GPR's higher accuracy. Similarly, for lattice parameter prediction, GPR demonstrated lower MSE (0.0023) and MAE (0.0396) values compared to RFR (MSE: 0.0025, MAE: 0.0423), with a slightly higher R-Square value (GPR: 0.9768, RFR: 0.9744), suggesting better overall goodness of fit. This study highlights the success of both RFR and GPR models in predicting lattice parameters and formation energy. It emphasises the potential of machine learning to efficiently and cost-effectively explore significant phase spaces in the search for new materials.
Keywords: Formation Energy, Lattice parameter, Machine learning, Heusler compound, Random Forest Regression (RFR), Gaussian Process Regression (GPR), R-squared (coefficient of correlation), Mean square error (MSE), Mean absolute error (MAE).
|
||||||
|
Date/Time: |
Abstract Number: ANPA2024-N00032 Presenting Author: Gokarna Banjade Presenter's Affiliation: Patan multiple campus Title: PREDICTION OF NEPAL'S 2021 EARTHQUAKE TREND USING SCALED INPUT IN A 4-LAYER LSTM MODEL. Location: Poster Presentation Show/Hide Abstract Earthquake prediction is a critical field of study in Nepal due to its susceptibility to seismic activity caused by the energy released during plate movements. In this work, we use the LSTM model and advanced computer programs to analyze earthquake data and find trends. We utilize a deep neural network approach to enhance prediction accuracy. The dataset used from May 4, 1964, to May 23, 2023, was obtained from the USGS (United States Geological Survey) catalog of magnitude Mw ≥ 4. In this work, we aimed to predict earthquakes and trends around the year 2021 using data from a series of past earthquakes. A dataset of 1094 data points was used,with 88 for training and 1020 for testing. The primary focus was on earthquake event 1040, which occurred on May 18, 2021. The trained model is then used to predict the future trend of
earthquakes. We consider four features like magnitude, time interval (in days), longitude, and latitude, and we find the best-fit model from the Kerus tuner-coded Python program. The program is executed multiple times, with each run corresponding to a specific number of epochs ranging from 100 to 1500. We observed that training the model with a larger number of epochs improved the
prediction accuracy for the longitude, latitude, time interval, and magnitude features. Specifically, we observed that during epoch 1000, the prediction for longitude, latitude, time interval, and
magnitude of observed data tends to match with test data and is better as compared with other epochs. Reviewing the graph and its data, it becomes clear that errors are significantly lower during
epoch 1000, as supported by the analysis of mean square error and mean absolute error.
|
||||||
|
Date/Time: |
Abstract Number: ANPA2024-N00031 Presenting Author: Ambika Shahi Presenter's Affiliation: Central Department of Physics, Tribhuvan University, Kirtipur 44613, Kathmandu, Nepal Title: Prediction of Lattice Parameter and Formation Energy of Cubic Semi-Heusler Alloys Location: Poster Presentation Show/Hide Abstract Materials advancement is the propelling forces for present-day science and technology. In recent years, machine learning (ML) models are becoming popular in the field of materials research due to their remarkable precision at low computational cost. Notably, semi-Heusler alloys are getting attention due to their applications in thermoelectric, topological, shape memory as well as spintronics. In this work, we apply ML models to predict the lattice parameter and formation energies of selected semi-Heusler alloys XYZ (where X and Y are transition elements, and Z is a main group element) using dataset sources from materials project database. We predict the lattice parameters of 145 cubic semi-Heusler alloys using 24 features through regression models, specifically, Random Forest Regression (RFR) and Gaussian Process Regression (GPR). Subsequently, the labelled datasets were split into training and testing sets in the ratio of 80:20 for both the lattice parameter and formation energy prediction. A feature versus contribution barplot was created to show the contribution level of each parameter for predicting results. Assessing the model’s performance, we computed R-square, mean square error, and mean absolute error for RFR (GPR) models. Our findings suggest that GPR out performs RFR in predicting the lattice parameter, while RFR predicts the formation energy of cubic semi-Heusler alloys better. Furthermore, a heatmap shows the correlation between the actual data and the predicted data. This study demonstrates the developing predictive models using machine learning in materials science which enables deeper understanding of lattice parameter and formation energies.
|
||||||
|
Date/Time: |
Abstract Number: ANPA2024-N00030 Presenting Author: Aatiz Ghimire Presenter's Affiliation: Tribhuvan University Title: Deep Learning Technique for Forest Fire Detection using Satellite Imagery Location: Poster Presentation Show/Hide Abstract Forest fire is one of the major causes of air pollution in Nepal, which also an fueling agent to climate change. To detect and identify forest fire is not an easy task in dense and isolated forest. To overcome this problem, we suggest a deep learning-based model to detect forest fire using satellite imagery.
The satellite image in this research is extracted from different opensource satellite data products such as Sentinenal-2 and Landsat Satellites. These data products contain multiple hyper-spectral band images, where we require only Red, Blue and Green Band Images. These images are then converted into True color Images. The Images containing the Forest Fire are then annotated manually with the help of label studio. This data is passed to Convolutional Neural Network to learn the forest fire features within the Satellite Imagery. This will detect fire, smoke and burnt areas in this Satellite Imagery. The dataset made for this purpose is split into 80-20 and made validation split to check the accuracy and performance of the model.
|
||||||
|
Date/Time: |
Abstract Number: ANPA2024-N00034 Presenting Author: Pratiksha Khanal Presenter's Affiliation: Central Department of Physics, Tribhuvan University, Kirtipur 44613, Kathmandu, Nepal Title: Predicting the Lattice Parameters of Transition Metal Halides - MXn via Machine Learning Location: Poster Presentation Show/Hide Abstract Transition metal halides have been studied extensively because of their wide range of tunable electrical, magnetic, and topological properties that come from their unique crystal structure. Knowing lattice parameters, a building block of every crystal structure, enriches filtering materials properties like structure stability, electronic structure, magnetic ordering, and thus materials discovery. Here, we present machine learning models: Random Forest Regression (RFR) and Gradient Boosting Regression (GBR) for predicting lattice parameters (a,b,c) of transition metal halides - Mxn with 238 samples based on the properties of their constituent atoms. The metrics R2 score and mean absolute error (MAE) are used for validation of model performance yielding the overall values 0.80±0.09(0.82±0.14) and 1.33±0.64(1.12±0.35) using RFR(GBR) respectively. We have observed that the predictive strengths of both models are quite comparable but as a whole, GBR model is found to be more favorable than RFR. These findings are expected to be useful to experimentalist and theorist in characterizing and studying various properties of Mxn system.
|
||||||
|
Date/Time: |
Abstract Number: ANPA2024-N00035 Presenting Author: Pratima Khadka Presenter's Affiliation: Central Department of Physics, Tribhuvan University, Kirtipur 44613 Kathmandu ,Nepal Title: Machine Learning Driven Prediction of Formation Energy and Band Gap of AxMyM'zO6 Oxides Location: Poster Presentation Show/Hide Abstract Machine learning (ML) is an efficient method in the discovery of new materials due to its lower computational costs as compared to conventional density functional theory (DFT) and laboratory trials. In this work, we are using ML approach to predict the formation energy and band gap of AxMyM'zO6 (x=1,2,3 & y,z=1,2 ; y+z=3) perovskite material. A dataset of 350 compounds is collected with following features: lattice parameter, band gap, formation energy, total magnetization, energy above hull, density, and periodic table properties. We are using four ML models: Random forest (RF) regressor, Gradient boosting regressor (GBR), Support vector regression (SVR) and CatBoost regressor (CBR). Root mean squared error (RMSE) and R-squared value are evaluated to determine the performance of different models. The contribution of different features for formation energy and band gap prediction are also plotted from RF model. It is found that CBR model has achieved maximum performance in prediction with R-squared and RMSE: (0.83 and 0.41) and (0.44 and 0.69) for formation energy and band gap respectively. The properties of materials depends significantly on the magnitude of band gap. Formation energy is a determining factor of material's stability and synthesizability. So, ML can be used to assist DFT calculations for material discovery and its applications.
Keywords: Machine learning, Formation energy, Band gap, RF, CBR.
|
||||||
|
Date/Time: |
Abstract Number: ANPA2024-N00049 Presenting Author: Madan Lamichhane, PhD (Invited) Presenter's Affiliation: FICO Title: Physicist's Journey to Machine Learning Fraud Prevention Scientist Location: Virtual Presentation Show/Hide Abstract The transition from a career in physics to one focused on machine learning for fraud prevention in financial systems represents a compelling narrative of interdisciplinary exploration and professional evolution. This abstract offers insights into this journey, highlighting the unique perspectives, challenges, and opportunities that physicists encounter when venturing into the realm of solving the real-world problem of fraud prevention through the machine learning model.
We physicists possess a diverse skill set encompassing advanced mathematical techniques, computational modeling, and a deep understanding of complex systems dynamics. As they pivot towards addressing the intricate challenges of fraud detection and prevention, physicists leverage their analytical prowess to uncover patterns, anomalies, and underlying structures within vast datasets of financial transactions.
This also includes the methodologies that physicists can employ for transitioning to machine learning scientists, emphasizing integrating physics principles into algorithmic design and model development. Drawing parallels between concepts in statistical mechanics, network theory, and quantum mechanics, physicists can craft innovative approaches to identify fraudulent activities with heightened accuracy and efficiency.
Furthermore, this also explores the pivotal role of interdisciplinary collaboration in facilitating this career transition. Physicists engaging with data science, cybersecurity, and finance experts should gain domain-specific knowledge and refine their understanding of fraud dynamics in financial ecosystems. Through collaborative efforts, physicists can enhance their ability to translate theoretical insights into practical, real-world solutions that mitigate risks and safeguard financial institutions and consumers.
This also reflects on the personal and professional growth experienced by physicists navigating this transition. From acquiring new technical skills and domain expertise to embracing challenges and cultivating adaptability, the journey from physicist to machine learning fraud prevention scientist fosters a spirit of continual learning and innovation.
In conclusion, this underscores the significance of interdisciplinary collaboration and the potential for cross-pollinating ideas between physics and machine learning in addressing real-world challenges such as fraud prevention. By embarking on this career transition, physicists can contribute their unique perspectives and analytical rigor to advancing the state-of-the-art in fraud detection and prevention, ultimately strengthening the resilience of financial systems in an increasingly digital age.
|
||||||
|
Date/Time: |
Abstract Number: ANPA2024-N00050 Presenting Author: Dipak Rimal and Puskar Chapagain Presenter's Affiliation: Benchmark Education Company, NewYork, NY, 10801 and bSouthern Arkansas University, Magnolia, AR, 71753 Title: Can AI Solve Physics Problems? Evaluating Efficacy of AI Models in Solving Higher Secondary Physics Exam Problems: A Comparative Study Location: Virtual Presentation Show/Hide Abstract The rapid advancement in the field of Generative Artificial Intelligence (AI) powered by large language models (LLMs) has expanded their application in diverse technical fields. While these models excel in various complex natural language processing tasks, they still lack the robust reasoning ability required to solve advanced mathematics and physics problems. Despite their current limitation, their adoption in higher education in physics and other technical fields is rising. In this project, we assessed the ability of various state-of-the-art proprietary and open-source AI models to solve Physics exam problems, specifically the board exam of class 12 of HSEB of Nepal. To achieve the goal, we employed a zero-shot prompting technique to generate solutions from the different LLMs and evaluated these solutions by the subject matter experts. In this talk, we will provide a brief overview of the underlying architecture of these models and compare the performance of different models as evaluated by human experts. Additionally, we will also discuss the risks and benefits of their use in higher education.
|
||||||
|
Date/Time: |
Abstract Number: ANPA2024-N00051 Presenting Author: Prajwol Lamichhane Presenter's Affiliation: University of North Florida Title: Comparative Analysis of ChatGPT, Bard, and LLaMA on Medical Prompts using the MedQuAD Dataset Location: Virtual Presentation Show/Hide Abstract This research conducts a comparative analysis of various AI language models—ChatGPT, Bard, and LLaMA—focusing on their responses to medical prompts related to conditions such as Cancer, Senior Health, Bodily Pain, and Cardiovascular diseases. The study aims to evaluate the accuracy, comprehensiveness, and appropriateness of the generated responses using the MedQuAD dataset. A standardized set of medical prompts is used to query each model, and responses are analyzed both qualitatively and quantitatively using syntactic and semantic similarities against manually curated answers. Additionally, the generative factor of these AI models is tested by identifying the most frequently used keywords in their responses. The findings highlight the strengths and weaknesses of each model in handling complex medical information and offer insights into their potential utility and limitations in the healthcare domain. This project is crucial for understanding how AI can be leveraged to support healthcare professionals and improve patient outcomes in real-world scenarios.
|
||||||