Theodoros Panagiotakopoulos

Orlando, Florida, USA, +1 (321) 202-3216
Theodoros.Panagiotakopoulos@ucf.edu , teosfp@hotmail.com , Download Resume

I am Theodoros Panagiotakopoulos, an experienced Ph.D. candidate in Computational Physics, with a deep specialization in modeling for Computational Lithography and Material Design. My work is driven by a passion for exploring the intersection of AI/ML for Modeling Engineering, Material Design, and Electrochemistry. Driven by a commitment to innovation and a proven track record of translating scientific concepts into Machine Learning problems, I am ready to be your trusted partner in advancing technological frontiers.

At the University of Central Florida, my Ph.D. research focuses on the cutting-edge approach of using Machine Learning to revolutionize nanotechnology and material science. I'm developing innovative Neural Network Potentials to accelerate Molecular Dynamics and Kinetic Monte Carlo simulations. My expertise in Quantitative Analysis, Data Science, Computational Physics, and modeling, and advanced Neural Networks enable me to push the boundaries of these fields.

My passion for AI reaches beyond conventional applications. Throughout my Ph.D., I am driven to apply its potential to real-world challenges, especially environmental ones. I am currently exploring AI-driven approaches for designing catalysts that optimize CO₂ reduction and developing algorithms to reduce the computational cost of Density Functional Theory (DFT) calculations. This dedication originates from my firm belief that technology can be a powerful tool for sustainability. By making these complex simulations more efficient, I aim to pave the way for faster discoveries in material science with a smaller environmental footprint.

In collaboration with the Department of Statistics, we are developing Graphical Neural Networks (GNNs), a cutting-edge AI tool aimed at transforming complex data into clear understanding. By using adaptable constraints, we can extract the most important information, simplifying the analysis process and highlighting key features for scientists and engineers in materials science, energy, and R&D. Utilizing GNNs enables us to understand data faster, make informed decisions, and design new, cutting-edge algorithms for simulation, ultimately conserving computational resources. Our goal is to apply this model in material design, significantly reducing the need for time-intensive simulations with limited atom numbers.

My comprehensive skill set includes quarantine analysis, Statistical Data Science, Computational Physics, Computational Modeling, Predictive Analytics, and extensive experience in Neural Networks, including Deep Learning and GANs. Adept at time series analysis and forecasting using AR, ARIMA, and deep learning techniques, I bring a multifaceted approach to problem-solving.

I hold a Master's degree from the National and Kapodistrian University of Athens, where I specialized in Deep Learning and Data Analytics. My master's program focused on the application of these advanced techniques to the detection of supersymmetry and dark matter, providing a solid foundation for my current research endeavors in Computational Material Science.

Beyond my research pursuits, I have served as an accomplished Teaching Assistant at the University of Central Florida and at the National and Kapodistrian University of Athens. I take great pride in guiding and shaping the next generation of minds in both graduate and undergraduate courses.

As a Modeling Engineer and Computational Physicist I am uniquely positioned to contribute transformative solutions to organizations seeking to elevate their technological capabilities. Collaborate with me to pioneer advancements in AI and ML. Together, we can harness technology's boundless potential for a more innovative future.

Experience

Modeling Product Engineer, ASML, Silicon Valley

I worked with the Tachyon software and studied geometrical corner rounding. I personally designed FEM+ simulations and calculated the total run time, memory usage, and grid dependence. By performing a runtime breakdown, I was able to calculate the corner rounding run time, the render area and render edge time, as well as the EM3D time. For the contour-to-contour difference, I designed LMC+ simulations and used the MPC layer as the golden rule for my comparison. After discussing the results with R&D, this update was included in the latest Tachyon release and introduced to one of our biggest customers. I found that by using an updated geometrical corner rounding value, we could reduce significantly the computational time and corresponding memory usage by. I wrote my own Python and C++ code to extract data from large datasets, transformed them into libraries, and shared them with the company, where they were adopted by the product engineering group. Additionally, I identified a bug in the Tachyon software that was causing asymmetric jogs after applying the geometrical corner rounding algorithm. I proposed a solution to the research and development team, which they subsequently adopted.

I worked on transition cross coefficient (TCC) optimization. I designed rigorous M3D simulations using Tachyon software, specifically FEM+. I utilized a variety of settings to calculate the optimal TCC number that yields the lowest aerial image critical dimension (AI_CD) difference between the baseline model and models with fewer TCC numbers. For this purpose, I used different settings and mask types, such as high numerical aperture with a low refraction index mask, high numerical aperture with a binary mask (with a large refractive index contrast, making the brighter side brighter and the dark side darker), low numerical aperture with a low refraction index, and low numerical aperture with a binary mask, each with different mask patterns. These patterns included one-dimensional patterns, two-dimensional patterns, circular patterns, elliptical patterns, and polygonal patterns. I calculated the optimal TCC number, which reduced significantly the runtime and the corresponding memory usage, while maintaining accuracy in the aerial image within a specification of 0.1 nm and causing a shift in the x and y directions of less than 0.01 nm. After discussing the results with the research and development team, we included this update in the next Tachyon software release, and the result was also adopted by our customer. Additionally, I wrote my own Python and C++ code to extract and process data from the simulations, using a completely different approach from my previous project. I developed custom scripts to automate the data extraction, filtering, and analysis processes, allowing for efficient handling of large datasets.

I also enhanced the API library we use for Tachyon software, based on ASML standards. I located and fixed several bugs, and updated the library with new functionalities that can be used to build simulations faster and obtain results more quickly. These improvements significantly enhanced performance and reliability for FEM+ applications. With dedication and attention to detail, I ensured that the API library is now more robust and efficient, supporting our team's ability to deliver high-quality simulations.

Research Assistant, University of Central Florida

Artificial Inteligence, DOE - NSF Grant

Introduced a second generation neural-network representation that is several orders of magnitude faster than Density Functional Theory (DFT). This method, relies on local properties and is not taking into account global changes in the electronic structure. This representation gives the energy and forces as a function of all atomic positions in systems of any size. An atomic neural network uses the local chemical environment to determine each atom’s energy. A feed-forward neural network, in which information only passes in one direction toward the output layer, is the sort of neural network that is used for fitting energy. There are two layers: an input layer that feeds the network with the respective atomic positions and an output layer that holds the atomic potential energy. The regression is carried out by a number of intermediary hidden layers, with the number of layers and the number of neurons in each layer being empirically optimized for the particular application. This accomplishment played a pivotal role in securing NSF funding .

Developed a general solution for the limitations of current Machine Learning potentials by introducing a fourth-generation Neural Network, which is applicable to long-range charge transfer and multiple charge states . It consists of highly accurate short-range atomic energies similar to those used in second generation neural networks potentials and charges determined from a charge equilibration method relying on electronegativities. Both, the short-range atomic energies as well as the electronegativities are expressed by atomic neural networks as a function of the chemical environments. For all these systems we demonstrate that fourth generation neural network potentials trained to DFT data are able to provide reliable energies, forces and charges in excellent agreement with electronic structure calculations. We show that previous generations of neural network potentials, which are unable to take distant structural changes into account, yield inaccurate energies and forces, which are correctly resolved by this neural network potential. These results apply to other types of Machine Learning potentials. This potential was adopted by our data science group, accelerating computational calculations.

Worked with the Department of Statistics in order to research is and investigate how graph convolutional neural networks can enhance the accuracy of predictions.The accuracy of predictions on a truncated dataset is examined in our experiments. The prediction made by averaging the parameters of two truncated datasets is then analyzed to determine if performance improvement can be observed. This is done by applying the methodological framework of the Simplified Graph Convolutional Neural Network(SGC) to Cora dataset. The impact of node removal on prediction accuracy is investigated by selecting nodes with three distinct patterns. Insights gained from these experiments shed light on which node patterns are crucial for optimizing the performance of SGC. If node removal is focused on a localized selection, the results demonstrate that SGC effectively utilizes the local information of network nodes. In this case, the applied method of improving the SGC’s accuracy, has better performance compared to a case where the nodes were randomly dropped. This facilitated collaboration between the Department of Statistics and Physics.

Data Modeling and Simulations, DOE Grant

Developed innovative numerical methods and algorithms for chemical potential calculations of metal on semiconductor junctions. The model successfully predicts the formation of metallic clusters on the semiconductor surface, a result experimentally confirmed by scanning tunneling microscopy (STM) experiments at UC Davis. Moreover, the island areas grew linearly with time, exhibiting collective diffusion, and their total growth rate was inversely related to the temperature. Density Functional Theory simulations of the chemical potential and binding sites of the Pb/Ge(111) system were used to explain this nonclassical behavior of the system. This led to a PRL article in collaboration with UC Davis.

Introduced a novel method to model the electrolyte at the electrochemical interface using initio simulations of charge transfer processes at surfaces. We presented a simple capacitor model of the interface that illuminates how to circumvent i) required large cell heights to reach convergence, which is a serious computational cost ii) the costly iterations calculations of reaction energetics to tune the surface charge to the desired potential. We derived a correction to the energy for finite cell heights to obtain the large cell energies at no additional computational expense. We furthermore demonstrated that the reaction energetics determined at constant charge are easily mapped to those at constant potential, which eliminates the need to apply iterative schemes to tune the system to a constant potential.

Investigated the cation effect using small quaternary ammonium cations with different sizes and symmetries. For Bi-catalyzed CO₂ Reduction Reaction ( CO₂ RR) that produces CO and formate, density functional theory (DFT) calculations and ab initio molecular dynamics (AIMD) simulations were performed to examine the effect of asymmetric cations (such as CH₃NH₃⁺ ) on CO₂ adsorption and activation. Experimental studies confirm that the asymmetric cations have a stronger promotional effect on the CO2RR activity than the symmetric ones, which is attributed to the asymmetric and weak hydration shell that stabilizes adsorbed CO₂ . Moreover, by comparing the cations with different sizes ( NH₄⁺ , Me₄N⁺ Et₄N⁺), we observed a notable effect of the cation size on CO production activity, with a negligible im- pact on formate production. Our work further elucidates the critical effects of cation symmetry and size on CO₂RR and suggests a method to improve electrocatalysis with optimized electrolytes. This led to the publication of a paper

Studied and reported the effect of creating an interface between a semiconducting polyaniline polymer or a polar poly-D-lysine molecular film and one of two valence tautomeric complexes, i.e., [CoIII(SQ)(Cat)(4-CN-py)2] ↔ [CoII(SQ)2(4-CN-py)2] and [CoIII(SQ)(Cat)(3-tpp)2] ↔ [CoII(SQ)2(3-tpp)2] . The electronic transitions and orbitals are identified using X-ray photoemission, X-ray absorption, inverse photoemission, and optical absorption spectroscopy measurements that are guided by density functional theory. Except for slightly modified binding energies and shifted orbital levels, the choice of the underlying substrate layer has little effect on the electronic structure. A prominent unoccupied ligand-to-metal charge transfer state exists in [CoIII(SQ)(Cat)(3-tpp)2] ↔ [CoII(SQ)2(3-tpp)2] that is virtually insensitive to the interface between the polymer and tautomeric complexes in the Co(II) high-spin state. This led to the publication of a paper.

Assist with compiling, debugging and optimization, of codes used in computational Material Science such as VASP, VASPsol and Quantum Espresso. Perform, analyse and summarize validation simulations on high-performance computer (HPC) systems running Linux. Assist in the implementation, development, and improvement of application software and methods that can be utilized in analyzing and interpreting data in the physical sciences. Four years active development of codes used in HPC.

Teaching - Consulting, UCF Funded

Recently started coaching and supervising new coming graduate students to acclimate themselves within the group and execute their research project. Provided computational help, assisted with class selection discussed with the advisor and group senior scientists about their research future steps. Keep abreast of new developments in Computation Material Science and be pro-active in introducing them to new graduate students. Help senior scientists with grant proposals by contributing sections of the proposal that describe the interplay between their research and high-end computing resources.

Specialized physics lab instructor, focusing on Machine Learning and data science. Led physics labs by incorporating comprehensive lessons on analyzing and applying simple Machine Learning models, while adeptly navigating artificial data landscapes derived from simulations I designed, as well as real datasets from our laboratory. Demonstrated a seamless fusion of physics and data science.

Guided undergraduates in mastering the intricacies of statistical data analysis, emphasizing the importance of data preparation for effective application of Machine Learning algorithms. Implemented feature engineering techniques, involving meticulous data cleaning and transformation, to enhance the quality and relevance of datasets

Research Assistant, National and Kapodistrian University of Athens

Deep Learning, NKUA Funded

Engaged in the cutting-edge development of a sophisticated Machine Learning Approach for Dark Matter Particle Identification, adeptly navigating the challenges posed by extremely low temperatures with unwavering precision and ingenious solutions. The model accurately predicts the origin of dark matter from the LSP.

Conducted immersive physics labs for undergraduates, delving into the intricacies of statistical data analysis and cultivating a deep understanding of the art of data preparation for the seamless application of advanced Machine Learning algorithms. Simulations were developed by our simulation group, providing us with a vast pool of artificial data for cleaning and training purposes

Education

University of Central Florida

PhD

Computational Physics

GPA: 4/4

August 2019 - Present

National and Kapodistrian University of Athens

Master's degree

Computational Physics

Grade: 9.2/10

My master's integrates deep learning methodologies to unravel the mysteries of supersymmetry and dark matter. The Standard Model, foundational to modern particle physics, falls short in addressing the hierarchy problem—the disparity between electroweak and gravitational forces. Supersymmetry, a hypothetical extension of the Standard Model, emerges as a compelling solution to this conundrum. Central to my research is the exploration of the Lightest Supersymmetric Particle (LSP), a theoretical entity holding profound implications for dark matter. Employing advanced deep learning techniques, I meticulously calculate the cross-section for the scattering of dark matter particles from ordinary matter, illuminating the elusive nature of dark matter. My investigation extends to both direct and indirect methods of dark matter detection, involving a thorough exploration of experimental setups and rigorous comparisons between theoretical predictions and observational astronomy data. Utilizing data science methodologies, I validate the robustness of the employed model, ensuring its alignment with real-world observations. In the realm of direct dark matter detection, I developed and trained a deep neural network specifically designed to identify the cross-section of dark matter particles, utilizing collected data from Fermi-LAT. The intricate workings of this neural network enhance the precision and reliability of our understanding of dark matter interactions. Moreover, my exploration extends to indirect methods of dark matter detection, including scenarios of annihilated dark matter particles and the observation of neutrinos and photons. In this context, another neural network I developed plays a crucial role, confirming the exclusive origins of neutrinos from specific particles. Throughout this scientific odyssey, these deep neural networks stand as pivotal tools within the overarching data science framework, refining and validating theoretical predictions against observational data from the vast expanse of the cosmos. My research thus culminates in a comprehensive understanding of the intricate interplay between supersymmetry and dark matter, with data science emerging as the indispensable tool for unraveling these cosmic enigmas.

October 2017 - May 2019

National and Kapodistrian University of Athens

Bachelor's degree

Physics

It was based on research paper that has been produced by other reinformation searchers. I had to study it, reproduce it and piece it an original presentation. According to my undergraduate thesis, reference is made not only to the decomposition reactions of π ⁰, but also to the probability that occurs in each of them. A study is also made to a detailed analysis of the structure and mode of operation of the Tevatron accelerator and of the CDF (Collider Detector at Fermilab), SVXII, ISL (Silicon Layers) and COT (Central Outer Tracker) detectors, which are the trace detection systems and are designed to detect the charged particle trajectories. Next, the conversion of γ → e ^- e⁺ to Tevatron is studied to reduce the uncertainty in the initial conversion probability. The above procedure is done by creating an algorithm. From the reconstruction of the mass of π⁰ and from the combination of the four trajectories, I find the cleavage position of π⁰ . Then we analyze the possibility of using the space variables in the decomposition of π⁰ while analyzing the topology of both normal and Daliz π ⁰. After the study we come to the selection of the appropriate variables which are defined as firstpoint and minpoint, where Monte Carlo technic plaid a crucial role. Finally, from the diagram that emerged from the breakpoints of π⁰ , we proceeded to identify the ISL, COT and SVXII detectors.

October 2011 - May 2017

Theodoros Panagiotakopoulos

Experience

Modeling Product Engineer, ASML, Silicon Valley

Research Assistant, University of Central Florida

Research Assistant, National and Kapodistrian University of Athens

Education

University of Central Florida

National and Kapodistrian University of Athens

National and Kapodistrian University of Athens

Material Science Skills

Data Science Skills

Coding Languages

Commonly used Libraries

Management and Communication Skills

Awards and Fellowships

Conferences

Interests

Selected Publications