𦫠I'm a Research Associate in Prof. Sam Peng's research group with research interests span physics, materials, and machine learning.
π Experimental science. Observe, record, and describe natural phenomena, summarize and verify laws.
- Galileo, Newton's laws, Mendel's Pea Experiment: The Hypothetical Deduction Method
π‘ Theoretical science. Based on mathematical analysis and theoretical derivation.
- Maxwell's equations, general relativity, quantum mechanics. Recover the mathematical formula and abstract physics theory from limited experimental data(e.g. spectrum) and thought experiment.
π» Computational science. Based on the natural world, we create the mathematical world, and now we build the numerical computation world.
Simulated nuclear tests, computational materials, etc. Data sources including experimental measurement, empirical methods, numerical solution from extensively verified physical equations.
π₯ AI4science
Combination of multiple academic disciplines results in the interdisciplinary and comprehensive methodology: computer-aided, data-driven, statistics-based, physics-informed method.
Use ML model to make predictions, and verify these conclusions by scientific approaches and supportive physics theorems.
"Scientific discovery in the age of artificial intelligence": https://www.nature.com/articles/s41586-023-06221-2
See more about the fourth paradigm: it's attractive and challengend to use symbolic regression to discover new physical laws
https://www.microsoft.com/en-us/research/wp-content/uploads/2009/10/Fourth_Paradigm.pdf
My methodological interest in this field lies in advancing multi-methods materials designs and the application of experimental and statistical analysis to DFT and machine learning. It is all about harnessing the power and beauty of physical laws and data.
Machine learning (ML) has shown great promise in advancing our understanding of various aspects of physics. By leveraging ML techniques, researchers can analyze large and complex datasets, identify patterns, and make predictions more efficiently than traditional computational methods. Here are some ways to use machine learning to investigate physics:
-
Data analysis and pattern recognition: Machine learning can help analyze large datasets generated from experiments or simulations in physics, identifying patterns and relationships that may be difficult to discern through manual analysis. Examples include detecting anomalies in particle collider data or identifying features in astronomical images.
-
Surrogate modeling: ML models can be trained to approximate complex physical models, reducing the computational cost of simulations. These surrogate models can be used to predict outcomes of physical systems more efficiently than traditional methods, enabling faster exploration of parameter spaces and optimization of system properties.
-
Model improvement and parameter estimation: Machine learning can be used to refine existing physical models by identifying discrepancies between model predictions and experimental data. ML can also assist in estimating model parameters or initial conditions, which can be challenging in complex systems with many degrees of freedom.
-
Inverse problems: Machine learning can be applied to solve inverse problems in physics, where the goal is to determine the underlying cause of an observed effect. Examples include determining the distribution of mass in a galaxy from gravitational lensing data or inferring material properties from scattering data in materials science.
-
Discovery of new physical laws: ML algorithms can be used to discover new laws and relationships in physics by analyzing data and identifying underlying patterns or correlations. For example, symbolic regression techniques can be employed to find functional forms that best describe the relationship between variables in a physical system.
-
Quantum computing and quantum information: ML techniques can be used to study quantum systems, such as optimizing quantum gate operations, designing error-correcting codes, or simulating quantum systems more efficiently.
To get started with using machine learning to investigate physics, follow these steps:
- Identify a problem or area of interest in physics that could benefit from machine learning techniques.
- Acquire a suitable dataset, either from experiments, simulations, or publicly available sources.
- Familiarize yourself with machine learning tools and libraries, such as TensorFlow, PyTorch, or scikit-learn, and choose the most appropriate ML algorithm for your problem.
- Preprocess and clean the data, ensuring it is properly formatted and suitable for training your ML model.
- Train and validate your ML model, fine-tuning hyperparameters to achieve optimal performance.
- Analyze and interpret the results, comparing your model's predictions with experimental or simulation data and assessing the model's performance.
π¨βπ» Currently, we delve deep into the fascinating realm of metallic chemical elements lanthanide:
Depart from lanthanides' abundant electronic structure in solid, we try to explore microscopic interactions(electron, photon, phonon, etc) and energy transfer mechanisms (ED, MD, MPR, up-conversion, cross-relaxation, etc) with Judd-Ofelt theory and DFT. Grounded in an extensive theoretical framework and calculated transition rates, we implement Monte Carlo model and machine learning algorithm to simulate the probabilistic, collective energy transfer behaviour among hundreds to millions of ions within single nanoparticles with diameters ranging from 4
π The overarching goal of these endeavors is to hack the energy transfer network to design custom-based nanomaterials: UCNP, perfect for their ultimate application: serving as long-term, multi-color optical probes in single-molecule tracking and imaging. Letβs push the boundaries of UCNPs research, illuminating paths for future innovations in spectroscopy and bio-imaging, and drive innovation together!
I seek to hack the energy transfer networks, build a useful computational model to provide insights and predictive support for the development of UCNPs. Here are some of my current most interested topics:
βοΈ How many term symbols
βοΈ How do we specify the existence of various energy transfer pathways in UCNP based on experimental observation and theoretical investigation beyond known categories such as electric-dipole induced radiative transition(ED), magnetic-dipole induced radiative transition(MD), multipolar interaction, multi-phonon relaxation, non-resonant donor-acceptor energy transfer process(ET)?
βοΈ How do we obtain the overlap integral
βοΈ What's the distinctions between MC and ODE? e.g. MC generates more blink curves than smooth population evolution curves generated by ODE; is this normal? How to handle with the interionic distance
βοΈ Doping percentages and MC model stochasticity: how do they relate? If large particles or high doping that result in more doped ions produce a more accurate simulation than small particles and low doping?
βοΈ Should we take into account the local site symmetry responsible for the selection rules at different lattice sites, as viewed from the perspective of a single ion? If not, is the selection rule applicable to every ion in the lattice environment?
βοΈ How many adjustable variables cound be optimizaed (lanthanides categories and percentages, crystal structure, particle size, laser power density and polarization, etc) that supprot precise tuning of the photonphysical properties and high energy transfer efficiency of UCNP?
βοΈ Could we explore the potential of correlating the Monte-Carlo model with neural networks due to their structural and dynamics similarities?
π οΈ I'm also interested in collecting some useful AI tools to improve productivity, they are my brainstorming partner and positive assistants with can-do attitude!