Machine Learning / Data science on non-Physics systems
DFT/MD Modeling and Simulations
PINNS and PIML
Dr. Marx Akl is a Computational Physicist, Principal Scientist, AI Research Advisor, and actively practicing AI scientist and engineer leading an AI-first research lab at the intersection of artificial intelligence, computational physics, materials science, engineering, and intelligent scientific systems. Every project within his lab—whether focused on physics, engineering, healthcare, molecular science, materials discovery, or predictive maintenance—is designed with a strong AI component at its core.
Dr. Akl holds a Ph.D. in Physics from Rensselaer Polytechnic Institute, where his dissertation focused on The Mechanical Response of Materials at the Nanoscale via Simulations. He also holds Masters' degrees in Physics and Mechanical Engineering from RPI and the University of Southern California. With over a decade of experience as a scientist, machine learning engineer, and AI systems builder, Dr. Akl brings together deep scientific expertise and modern AI engineering to solve complex real-world problems.
His current work spans large language models (LLMs), retrieval-augmented generation (RAG), agentic AI frameworks, computer vision, transformer architectures, neural networks and operators, multimodal learning, predictive maintenance, molecular machine learning, and AI-driven scientific discovery. He actively develops and advises projects involving many areas such as real-time object recognition for assistive wearable technology, transformer models for molecular representation learning, machine learning models for turbofan engine failure prediction and Remaining Useful Life estimation, autoencoder-based integration of single-cell imaging and sequencing data, and operator-learning frameworks for chaotic dynamical systems.
The lab integrates modern AI methods with traditional scientific computing techniques such as density functional theory (DFT), molecular dynamics (MD), simulation, spectral analysis, and physics-informed modeling. Rather than treating AI as a post-processing tool, Dr. Akl embeds AI directly into the research workflow—using it to accelerate discovery, reduce computational bottlenecks, uncover hidden patterns, guide hypothesis generation, and build intelligent systems capable of reasoning across structured data, unstructured documents, images, sensor streams, and scientific simulations.
Dr. Akl’s research includes AI-powered wearable systems such as Guide Glass for visually impaired navigation, transformer-based models such as DeepBERTa for molecular property prediction, hybrid ML frameworks for rapid band gap prediction in materials science, and advanced deep learning approaches for cancer data analysis and chaotic system forecasting. In parallel, his applied AI engineering work includes modern enterprise AI architectures such as LLM-powered RAG systems and agentic AI frameworks, including Azure-based agentic systems that connect models with tools, data sources, workflows, and real-world applications.
What distinguishes Dr. Akl’s work is his ability to bridge physics-driven scientific reasoning with cutting-edge AI engineering. His lab operates at the convergence of generative AI, agentic AI, computer vision, scientific machine learning, materials informatics, intelligent simulation, and applied data science. The goal is not simply to use AI as a tool, but to build AI systems that can assist, reason, predict, discover, and act within complex scientific and engineering environments.
Dr. Akl’s research philosophy is simple: AI is not an add-on — it is the core engine driving modern scientific discovery.
Machine Learning in Materials Science/Physics: Dr. Akl leads pioneering research in applying machine learning techniques to enhance density functional theory (DFT) and Molecular Dynamics (MD) simulations, characterize material properties, construct precise energy band structures, and analyze intricate phonon dispersion curves.
Machine learning (ML) can be effectively integrated with Density Functional Theory (DFT) simulations to enhance the accuracy and efficiency of electronic structure calculations. Here are specific examples of how I plan to use ML techniques to various aspects of DFT simulations:
1. Predicting Total Energies:
- Challenge: Calculating total energies in DFT simulations can be computationally expensive.
- ML Application: Train a regression model (e.g., kernel ridge regression or neural networks) to predict total energies based on atomic configurations.
- Example: The ANI (Atomic Neural Network) potential is trained on DFT data to predict total energies, enabling fast and accurate energy evaluations for molecular dynamics simulations.
2. Accelerating Structural Relaxations:
- Challenge: Optimization of atomic positions in a material can be time-consuming.
- ML Application: Use ML models to predict forces acting on atoms, enabling accelerated structural relaxations.
- Example: The FCHL (Fingerprints for Chemical Learning) framework employs a neural network to predict forces and accelerates geometry optimizations in DFT calculations.
3. Basis Set Corrections:
- Challenge: DFT calculations often use finite basis sets, leading to systematic errors.
- ML Application: Train a model to correct basis set errors and improve the accuracy of DFT results.
- Example: The Gaussian Approximation Potentials (GAP) corrects basis set deficiencies by learning the correlation between DFT-calculated properties and atomic environments.
4. Band Gap Prediction:
- Challenge: Predicting electronic band gaps accurately is crucial for understanding a material's properties.
- ML Application: Develop ML models to predict band gaps based on electronic structure features.
- Example: The Materials Project uses ML algorithms to predict band gaps, providing efficient screening of materials for specific electronic properties.
5. Materials Discovery:
- Challenge: Identifying new materials with desired properties is time-consuming through traditional simulations.
- ML Application: Employ ML models to predict material properties and guide the search for novel materials.
- Example: The AFLOW-ML project combines DFT data with ML to accelerate the discovery of new materials with specific electronic, thermal, or mechanical properties.
6. Charge Density Mapping:
- Challenge: Visualizing charge density distributions can be computationally intensive.
- ML Application: Train ML models to predict charge density distributions based on atomic configurations.
- Example: ML-based approaches, like Deep Density Networks, offer fast and accurate predictions of charge density distributions, facilitating visualization and analysis.
Data Science / Machine Learning: We harness the power of Data Science and Machine Learning to delve deep into diverse datasets, extracting valuable insights and driving meaningful outcomes. Our expertise spans a wide range of applications, from predicting home prices in housing datasets to advancing critical healthcare solutions like Harmful Brain Activity Classification.
We leverage advanced machine learning models, including linear regression, decision trees, and ensemble methods, to analyze housing datasets. Our algorithms delve into factors such as square footage, location, and amenities to accurately predict home prices. Whether you're a homebuyer, seller, or real estate enthusiast, our insights empower you to make informed decisions in the dynamic real estate market.
In the realm of healthcare, we are committed to making a positive impact. Our Harmful Brain Activity Classification project employs sophisticated machine learning techniques, including deep learning neural networks and support vector machines. By analyzing electroencephalogram (EEG) data from critically ill patients, we aim to classify seizures and identify patterns of harmful brain activity. This groundbreaking work has the potential to enhance patient care and revolutionize how we monitor neurological health in clinical settings.
We showcase the versatility of data science and machine learning. From financial predictions to healthcare diagnostics, our projects highlight the transformative potential of leveraging data. Our commitment to innovation, coupled with a deep understanding of cutting-edge models like Random Forests, Gradient Boosting, and Convolutional Neural Networks, positions us at the forefront of data-driven exploration.
Join us on a journey of discovery where data science meets real-world challenges. Explore the endless possibilities of machine learning in shaping industries, enhancing decision-making, and driving progress. Whether you're a data enthusiast, industry professional, or someone curious about the intersection of data and innovation, our lab is your gateway to the exciting world of data-driven excellence.
Uncover insights, make informed decisions, and embrace the future with Dr. Akl’s lab. Together, let's turn data into action.
Material Science and Solid-state Physics: The group investigates the mechanical responses of nanoparticles under tension and compression, revealing trends and critical transitions that contribute to a deeper understanding of nanomechanical behavior. We also characterize materials to reveal their mechanical properties such as stiffness hardness surface and fracture energy viscosity strength hardness and many others. We inspect a multitude of factors that can influence this mechanical behavior such as size effect, temperature, simulation variables shear bands, shear and normal stresses, inherent defects and pre-existing cracks and how each affect material behaviour and properties.
The compression of nanoparticles is of fundamental importance both scientifically, and to various practical applications such as crack resistant paint, ceramic coating for solar cells, tribology, targeted drug delivery, biosensors, and many others. We embark on an extensive investigation of the response of a nanoparticle to external compression. We investigate how shear bands formations are size independent at the early stages of plastic deformation but become size dependent at later stages of plastic deformation. I published earlier a size-dependent brittle to ductile transition between and for binary metallic glass MG nanoparticle coupled with a Griffith based fracture model that calculates to the same order of magnitude. We also investigate the effect of surface tensile and compressive stress states on fracture. We can apply both states using thermally tempered nanoparticles in MD simulations as was done in the published work. We show that compressive surface stress strengthens the nanoparticle by inhibiting crack formation even with pre-existing cracks present. We show that tensile surface stress further weakens the material. Both states however still promoted shear deformation which remains an opened question for a project investigation. We demonstrate while tensile surface promotes both shear and fracture, compressive surface promotes shear and inhibits fracture. This was the case with same material indenter and a repulsive indenter. We map the force displacement plot features to fracture morphology of the nanoparticle as compression ensues. We confirm that pre-existing cracks give rise to fracture much faster, and sooner. Our compressive surface layer is able to mediate that effect and delay fracture and reduce its severity as can be further investigated using other materials and varied shell thickness and relaxation states.
Advanced Modeling and Simulation Techniques DFT and MD: Leveraging state-of-the-art tools such as LAMMPS, Quantum ESPRESSO, and more, Dr. Akl's team conducts simulations in areas ranging from molecular dynamics to quantum mechanics. Both DFT and MD simulations play crucial roles in investigating and calculating mechanical properties of materials. DFT provides a theoretical foundation by predicting electronic and structural properties that influence mechanical behavior, while MD simulations offer dynamic insights into how these materials respond to external forces and environmental conditions. When applied to exfoliation regimes, these simulations help elucidate the factors influencing the stability and ease of layer separation in layered materials, guiding the design of novel materials with tailored mechanical properties. By synergistically utilizing DFT and MD simulations, researchers can gain a comprehensive understanding of material behavior, leading to advancements in materials science and engineering.
Epitaxial methods exfoliation regime of thin film/substrate perovskites crystals:
We design model and run DFT calculations on heterogenous perovskite systems to investigate exfoliation. It is an invaluable mean of detaching epitaxial layers from substrates to produce membranes that are essential in various applications such as optoelectronics high speed computing and various others. We calculate all interface energies from which we devise an exfoliation regime that matches experimental results in every case. As PTO-STO has the weakest interface and BTO-STO has the strongest, toughest interface we construct charge density difference maps and Bader analysis of charge using DFT to explain it at the chemical level. We find in a prior work that the transfer of electrons is greater in the BTO-STO than PTO-STO by a factor of two. Much more remains to be investigated. We also investigate exfoliation using MD simulations. While interrogating the effect of interface strength and stressor thickness on exfoliation we attempt to show that the stronger the interface the more stressor thickness is needed for exfoliation. We find exfoliation even at the weakest interface is not possible without applying a minimal amount of stressor layer. We show that if a stressor thickness of is used in our model, at any interface interaction strength, , below 60%, and no matter whether there is a crack or not, and whatever the nature of the crack: an interface, film or substrate crack, the material eventually yields successful exfoliation at the interface. At above interface strength however, spalling always occurs instead no matter the location of the crack or its absence. This is worth investigating and is the subject of at least one project.
Furthermore we demonstrate and will further validate that the presence of a stressor is a necessary but not sufficient condition for exfoliation. We attempt to also relate stressor thickness and interface strength to the exfoliation event.
Graphing mediated Monte Carlo deposition simulations :
As successful peeling is contingent on defect free film-interface-substrate. We use MD to simulate and confirm that graphene nanopatterning allows for the great reduction of defects in freestanding single-crystalline membranes using deposition simulations of Ge on Si. We observe that thicker rigid masks of graphene give rise to defect formation. The deposition in the absence of graphene is ripe with dislocations. As we add a flexible graphene sheet (mask) the simulation is defect free and all dislocations and stacking faults disappear. If graphene was rigid, dislocations at the interface and stacking faults re-appear. We show that as graphene coverage increases the dislocation density is greatly reduced. We define and characterize a critical length, above which dislocations cease to form. But the upper limit of that range, if any is yet to be determined.
We also show as adhesion between graphene and Ge gets stronger, there is an increased dislocation formation, in both rigid and relaxed graphene cases. We plan to relate that directly to average stress, . The latter should increase as adhesion is stronger and dislocations should form. We find stripped graphene with same exposed area very unstable and experiences the largest stress/strain fields and hence the highest dislocation density of all simulations. The mechanism leading to that remains in question and would serve for a full project, possibly more depending on findings. Finally, we simulate and show the effectiveness of growing and harvesting multilayered epitaxial systems through multiple graphene layers. This results in layer-by layer peeling culminating in multiple free-standing membranes. It remains to be seen whether the increase in the number of layers will change the material, mechanical and physical properties and that is a topic for yet another research project.
Machine Learning Dynamics: Revolutionizing Classical Mechanics on Variable Friction Surfaces for Innovation and Impact
We seamlessly integrate classical mechanics with the power of machine learning. Our research focuses on unraveling the dynamics of moving objects on surfaces with varying friction coefficients. Through meticulous data collection using high-speed cameras, accelerometers, and force sensors, we capture the object's position, velocity, and acceleration as it moves on changing friction surfaces. Employing machine learning models, including linear regression, polynomial regression, and neural networks, we aim to predict and analyze the impact of friction on the trajectory of the moving object. Our objectives include understanding how different surfaces influence dynamics, predicting trajectories under varying friction conditions, and optimizing motion parameters for specific outcomes. Join us in exploring this unique intersection, where classical mechanics meets cutting-edge machine learning, unlocking new insights and applications in fields from robotics to transportation. Dr. Akl’s lab invites enthusiasts, students, and researchers to engage with our research findings, datasets, and methodologies, fostering a deeper understanding of motion dynamics and sparking innovation in physics and technology.
Solid State Physics/Material Science : New Method to measure Liquid Viscosity
Viscosity (η) measures the fluid’s resistance to flow . It is a very fundamental property of materials as it influences almost every physical property such as softening and melting etc. Prior work used the potential energy approach to find the viscosity at low stresses .
Future work: Very small stresses give rise to very small strain rates as well. As can be seen in Figure the plot for the logarithmic viscosity scales linearly with 𝜎2 at very low stresses and diverges at large stresses. Low strain rates are very hard to measure in an MD simulation. That’s because the change in the simulation box is very minute and hard to quantify. However, after shearing as shown in schematic Figure 14 every atom breaks bonds and forms other bonds and/or maintains certain bonds. The total number of events occurring at the atomic level is the sum of bond breaking and bond forming events. Since the overall strain has to be a total sum of those individual events, the hope is to be able to relate the total events to the strain rate/time as shown in schematic Figure.
To achieve this I wrote a Python code - 800 lines – that would perform two tasks: one it would generate the species dependent radial distribution function for the system at every frame so that the total number of nearest neighbors for every atom at every frame dumped can be discerned. Two it would track every atom throughout the duration of the simulation at every frame and calculates the new total number of neighbors as the atom moves to a new location and compares it to the initial neighbors i.e. at frame 0. It then calculates the total new events for the entire system. I used the code to plot these two tasks as shown in Figures below. Once we gather all this data and calibrate the system and test it at different dumps we can then scale the events with the strain rate and use that to calculate the viscosity at very small strain rates that would otherwise cannot be quantified accurately.
