Explore Projects

<p>Virtually all properties of interest arise from electronic interactions between the atoms of a crystal.</p> <p>The work I propose aims to develop a physics-informed graph convolutional neural network architecture, inspired by the <a href="https://en.wikipedia.org/wiki/Tight_binding" target="_blank">Tight-Binding model</a> formalism. This would help us initialise weights with information from the Slater-Koster tables made available as a priori, which would theoretically help derive a more realistic, accurate and data-efficient representation.</p> <p>In addition to this, I also aim to help design and develop essential parts of <a href="https://github.com/Chemellia" target="_blank">Chemellia</a>; primarily <a href="https://github.com/Chemellia/ChemistryFeaturization.jl" target="_blank">ChemistryFeaturization.jl</a>.</p>

<p>The goal of this project is to implement the <a href="https://deepmind.com/blog/article/muzero-mastering-go-chess-shogi-and-atari-without-rules" target="_blank">MuZero</a> algorithm in Julia, with the following expected benefits:</p> <ul> <li>attract people from the model-based RL community to Julia</li> <li>showcase Julia in an ambitious project that leverages its greatest strengths, and which could not be achieved using another language (with the same level of performances vs flexibility and simplicity)</li> <li>make another step towards a generic and flexible RL framework by implementing MuZero on top of <a href="https://github.com/jonathan-laurent/AlphaZero.jl" target="_blank">AlphaZero.jl</a>, reusing as much code as possible and creating new abstractions if necessary</li> </ul>

This project focused on making patient-level prediction (PLP) more practical within the JuliaHealth ecosystem by building some of the missing foundations around OMOP CDM data. The main challenges were around data preprocessing and consistent cohort construction, which are essential before any prediction pipeline can be reliable. To address this, the project delivered two key pieces: first, a standardized Tables.jl interface in HealthBase.jl to make OMOP CDM data easier to work with and preprocess; and second, a new package, OMOPCDMFeasibility.jl, which adds tools for checking cohort feasibility and data readiness. Both were designed to work hand-in-hand with existing JuliaHealth packages like OMOPCDMCohortCreator.jl and OHDSICohortExpressions.jl. Together, these contributions provide a more reproducible, end-to-end path for building PLP models directly on OMOP CDM datasets.

<p>In the last decades, Bayesian statistics has gained ground in the modelling of phenomena. Despite its advantages, to implement a Bayesian framework is still difficult and restricted because the available literature on Bayesian statistics generally focuses on mathematical formalism and requires greater statistical expertise than some other methods. For this reason, any tool or resource that facilitates the understanding of Bayesian statistics and its mathematical background will be very useful. Turing.jl is a high-performance probabilistic programming bayesian inference language inside Julia ecosystem. Inside Turing lang, MCMCChains.jl represents an implementation of Julia types for analyzing, storing and summarizing MCMC simulations and uses utility functions for diagnostics and results visualizations. This project comprises various tasks to improve different aspects of MCMCChains.jl such as plotting functionality and storage of MCMC sampling to achieve better heuristics.</p>

This project aims to develop a Julia module for simulating and optimizing Bell distillation circuits, addressing the challenge of selecting efficient entanglement purification strategies under varying noise and resource constraints. I will first construct a structured database of Bell distillation circuits with associated performance metrics. I will then design an optimization algorithm that selects the most suitable circuit based on inputs such as target fidelity, error rates, and resource availability. The project will include a simulation framework for evaluating circuit performance under configurable noise models. Deliverables include the circuit database, simulation tools, optimization algorithm, and a well-documented, user-friendly Julia product that integrates with existing quantum simulation workflows. This tool will enable researchers to systematically explore and deploy optimal distillation strategies, improving efficiency and accelerating experimentation in quantum computing.

<p>This is a proposal for the project <em>"Native Julia solvers for ordinary differential equations and algebraic differential equations"</em>. There is no solver for ordinary differential equation (ODE) boundary value problem (BVP) yet in the <em>JuliaDiffEq</em> organization. I am going to implement native Julia solvers for ODE BVP problem. There are a number for advantages of Julia native solvers over calling C or Fortran's existing libraries via Julia's FFI, including the ability to use a wide variety of types in Julia like <em>BigFloat</em> for arbitrary precision, easier for people to contribute and better integrate with other packages in Julia ecosystem. The goals for this project are to implement BVP related data structures, a shooting method and a collocation method. Together, this contribution will help make Julia a viable environment for researchers and scientists to simulate models from boundary value problems.</p>

Reachability analysis deals with uncertainty propagation in dynamical systems. Given an initial set, modeling the uncertainty on the initial state, it aims to answer the question "What is the set containing all possible final states?" This question is relevant in several domains such as safety-critical applications and systems verification. Traditional approaches have mainly concentrated to convex sets, which can lead to overestimation of the final state when this is actually non-convex (so-called wrapping effect). Recently, Sparse Polynomial Zonotopes (SPZ) have been introduced as a set representation for both convex and non-convex sets. SPZ have been shown to have a compact representation that allows efficient manipulation. The goal of this project is to implement Sparse Polynomial Zonotopes in LazySets.jl and the corresponding reachability algorithms in ReachabilityAnalysis.jl for the JuliaReach organization. The final result will be a state-of-the-art (SOTA) toolbox for reachability analysis with sparse polynomial zonotopes, together with systematic benchmarks comparing our Julia implementation both to different implementations of the same approach and other SOTA approaches in reachability analysis.

<p>Graphs are a ubiquitous and versatile data structure, which allow the representation of problems and systems across a vast array of domains like infrastructure networks, molecular models, language, and social interaction. Recent community work on graph computations has culminated in a standard API which casts graph computations as generalized linear algebra on sparse matrices. This enables performant, standardized graph algorithms which are easily differentiable by modern auto-diff tools. In this JSOC 2021 project I aim to provide automatic differentiation rules for all of the operations in the GraphBLAS C API which admit such rules. The SuiteSparse GraphBLAS (SS:GrB) implementation will be the primary target, and support for the several extensions to the specification in SS:GrB is also a goal. The primary application I am pursuing is an efficient implementation of graph machine learning in Julia, but it should be noted that graph computation and generalized sparse linear algebra have incredibly broad use cases including graph algorithms like breadth first search, and faster neural networks by taking advantage of sparsity and operators faster than typical multiplication.</p>

The hardware necessary to build networks, sensors and computers of quantum nature is at its infancy. Potential quantum memories, repeaters, and switches are being researched all over the world, and the quantum hardware race has started. Therefore, a need for efficient, functional, and accurate open-source quantum networking modelling tools has never been more imminent. Projects such as Julia’s QuantumClifford.jl enable researchers across the globe to research various aspects of quantum networking while anticipating the creation of the necessary hardware to produce laboratory experiments. They present cheap and efficient environments to start building potential switch, protocol, and network designs. In classical computing, if one wants to protect a bit against errors, it can often suffice to represent that bit by a logical bit by repeating the data bit. Nevertheless, in a quantum setting, that type measurement is much more complex. Quantum qubit theory along with quantum noise present one of the most challenging problems that the quantum information field must overcome to become scalable and profitable: How to account for the error correction, needed due to quantum noise, in quantum gates? Simulators, such as Julia's Clifford gate simulator, must mimic not only the behaviour of quantum noise, but also the effects of currently proposed quantum error correction codes. For this reason, this project proses a comprehensive library of frequently used quantum error correcting codes for Julia’s simulator of Quantum Clifford Circuits. These would include the Steane, Shor, and Toric codes, as well as other Surface and CSS quantum error correction codes.

<p><a href="https://github.com/0x0f0f0f/Metatheory.jl/" target="_blank">Metatheory.jl</a> is a general purpose metaprogramming and algebraic computation library for the Julia programming language, supporting the novel technique of e-graph term rewriting through equality saturation, first introduced in <a href="https://egraphs-good.github.io/" target="_blank">egg</a>. Our contribution allows performing general purpose symbolic mathematics, manipulation, optimization, synthesis or analysis of syntactically valid Julia expressions and user defined types, with a clean and concise programming interface, both during compilation or execution of programs. We have been currently experimenting with optimizing mathematical code and equational theorem proving strategies. Some of the proposed tasks for this project include:</p> <ul> <li>Experimenting with <a href="https://github.com/JuliaSymbolics/Symbolics.jl" target="_blank">Symbolics.jl</a> integration and optimization for symbolic mathematics.</li> <li>Experiments in optimizing lowered Julia IRs (possibly by statically overlaying methods with <a href="https://github.com/femtomc/Mixtape.jl" target="_blank">Mixtape.jl</a>), allowing users to define concise rewrite systems that directly optimize Julia code.</li> <li>Experimenting in encoding category theory axioms to efficiently model string diagrammatic equational reasoning, and to produce human readable equational proofs. Possibly, integrate this rewrite system with <a href="https://github.com/AlgebraicJulia/Catlab.jl" target="_blank">Catlab.jl</a></li> </ul>

This project focuses on improving the computational efficiency and biological realism of synaptic modeling in HNN-Core by refactoring how synapses are placed, simulated, and recorded within neuron models. Currently, synapses are instantiated at fixed locations across neuronal sections, including inactive ones, leading to unnecessary computational overhead and reduced clarity in simulation outputs. The proposed approach introduces parameterized synapse placement, ensuring that synapses are only created at locations where neurons receive thalamic or cortical input. The refactor will also restrict synaptic current recording to active synapses, reducing memory usage and improving interpretability of results. Additional enhancements include updating existing network models, ensuring backward compatibility, and developing validation strategies to maintain numerical and biophysical consistency. To further improve usability, the project will incorporate visualization tools for analyzing synaptic currents and placement patterns. Time permitting, biologically realistic variability in synapse positioning will be introduced to better reflect neural heterogeneity. While this work focuses on synaptic modeling, my recent contributions to HNN-Core have also involved improving testing infrastructure and code reliability, providing additional insight into maintaining correctness and stability in scientific simulation software. The outcome of this project will be a more efficient, interpretable, and scientifically robust synaptic modeling framework within HNN-Core.

<p>The Kuadrant Console Plugin provides a web interface for managing API gateway policies in OpenShift, but currently relies heavily on YAML editing for most policy types. While DNSPolicy and TLSPolicy already have user-friendly form-based interfaces with dual Form/YAML views, validation, and guided workflows, the remaining core policies (RateLimitPolicy, TokenRateLimitPolicy, AuthPolicy, Plan, and OIDC) still require users to manually write YAML. This creates a steep learning curve and error-prone configuration experience. This project aims to bring RateLimitPolicy, TokenRateLimitPolicy, AuthPolicy, Plan, and OIDC policies to feature parity with the existing DNS and TLS form implementations. Form designs will be provided by the Kuadrant team. The mentee will implement these designs as PatternFly-based form interfaces following the established patterns from the DNS and TLS policy forms. These forms will allow users to configure policies through validated form fields while maintaining the flexibility to switch to YAML view for advanced use cases. The forms must support both creation and editing of policies, include proper field validation, handle complex nested structures (such as rate limit configurations and authentication rules), and synchronize seamlessly between form and YAML representations using the same patterns already proven in the DNS and TLS implementations.</p><p><br></p><p>Expected Outcome:</p><p>&nbsp;- Form-based creation and editing interfaces for RateLimitPolicy, TokenRateLimitPolicy, AuthPolicy, Plan, and OIDC policies implemented using the same patterns, components, and structure as the existing DNSPolicy and TLSPolicy forms</p><p>&nbsp;- Dual view toggle (Form View / YAML View) with bidirectional synchronization using js-yaml for all five policy types</p><p>&nbsp;- Field validation following the established validation pattern covering required fields, numeric constraints, conditional dependencies, and Kubernetes resource naming conventions</p><p>&nbsp;- PatternFly component integration matching existing forms: expandable sections for complex nested configurations, validated text inputs, dropdowns for enum fields, and reuse of gateway selection components</p><p>&nbsp;- Policy-specific form fields for: rate limit units and counters (RateLimitPolicy), token-based rate limiting (TokenRateLimitPolicy), authentication strategies and credentials (AuthPolicy), plan tiers and quotas (Plan), and OIDC provider configurations (OIDC)</p><p>&nbsp;- Error handling using the existing error modal and inline validation message patterns</p><p>&nbsp;- Internationalization support for all form labels and validation messages using i18next following the existing localization structure</p><p>&nbsp;- Both create (`/~new`) and edit (`/:name/edit`) routes for each policy type matching the DNS/TLS routing pattern</p><p>&nbsp;- Unit and component tests covering form validation, YAML synchronization, and error states following the established testing patterns</p><p><br></p>

OGC APIs — Features, Records, EDR, and Processes — are powerful but inaccessible to non-developers. This project builds a formal MCP mapping specification that translates OGC API operations into structured tools any LLM can use, enabling urban planners and researchers to query geospatial data and chain complex workflows using plain English. Deliverables: a language-agnostic JSON mapping spec covering all four OGC API types, a Python reference implementation, and a browser-native showcase requiring no GIS expertise.

Since CVA6/Ariane is highly modular by design, it is relatively easy to develop and integrate multiple L1 caches. Additionally, because CVA6/Ariane is open source, users can develop L1 caches customized for their needs. The primary technological feature of HPDC is its L1 Dcache, which supports Multi-Requester, Multi-Issue, and Out-of-Order operations. This capability is especially valuable for HPC systems striving for high memory throughput. However, HPDC is still under development, and various functional enhancements are being implemented or planned to make it more useful for HPC systems by providing a variety of functions for diverse cores. In this year's GSoC, the objective is to first expand the functionality of the L1 cache for the CVA6/Ariane core, and then to make it scalable to accommodate many cores, laying the groundwork for making HPDC an even more significant entity for HPC systems. By exploring the appropriate scale and structure, it's possible to enhance HPDC's functions to be used as an instruction cache with scalability in mind. Furthermore, by connecting with OpenPiton, it becomes extremely easy to scale to large systems with many cores. This aligns with the goals of the GSoC project.

This project aims to enhance SU2's computational fluid dynamics (CFD) capabilities by optimizing the Neural Network implementation (MLPCpp) within its Non-Ideal Compressible Fluid Dynamics module. I will address key performance bottlenecks in the current system, including inefficient variable mapping, suboptimal network selection, sequential evaluation, and limited inverse regression approaches. My solution involves implementing hash maps for faster lookups, refactoring network selection, improving evaluation with SIMD instructions, batched inference, and improving inverse regression with literature-backed starting points. The project will be delivered through two main pull requests: the first focused on an improved data-driven tutorial with step-by-step guides and troubleshooting resources, and the second containing the optimized source code implementation and comprehensive benchmarking results, making data-driven approaches more accessible and efficient for all users.

This project addresses the computational cost of full Geant4 calorimeter shower simulation by constructing compact, readout-segmentation–independent point clouds from Geant4 steps for fast-simulation training. I implemented and evaluated three clustering methods—DBSCAN, CLUE-style approach, and an axis-aligned Grid Clustering approach—together with a validation suite (longitudinal/radial profiles, containment radii, raw moments, energy closure) and a side-by-side 3D viewer. Across parameter scans, DBSCAN and CLUE achieved compression but degraded radial structure (and DBSCAN showed mm-level longitudinal drift), whereas the Grid method preserved longitudinal and radial profiles while remaining fast and angle-robust.

<p>Machine Learning has the ability to gain information, process it and give a well-defined output to the end-user. Machine Learning algorithms can recognize patterns in behaviour and create their own logic. I will be applying Machine Learning algorithms and deep learning classification techniques to predict the onset of fever in patients. I plan to do the following over the summer. • Analyze the data and select the relevant physiological variables from the dataset. • Extract the features from the physiological variables. • Use different Machine Learning algorithms on the processed data. • Use deep learning techniques on the processed data.</p>

The Virtual Brain (TVB) is a scientific simulation platform that provides all means to generate, manipulate and visualize connectivity and network dynamics of the brain networks. Researchers in Computational Neuroscience use brain network models to understand the dynamic behavior of the healthy and diseased brain, as measured by various neuroimaging techniques such as fMRI, EEG, and MEG. TVB is currently the only neuroinformatics project providing a platform for researchers to work together on modeling studies using brain network models with realistic connectivity. The modeling of the simulations requires datasets generally hosted on data storage platforms like Zenodo, OpenSourceBrain, EBRAINS, etc. One has to download the dataset manually, unzip it and then use it inside the web GUI. Instead of manually downloading and unzipping the data, the project's goal is to have a dedicated framework to manage the downloading and unzipping the dataset from remote sources. Right now, we are focusing on the Zenodo platform, but there are other platforms like OpenSourceBrain and EBRAINS for which we would like to extend the downloading functionality.

<p>OpenWorm is an attempt to build a biophysical simulation of the model organism C. elegans via assimilation of published data from a range of sources. Publication of biological data is often weakly structured and has no standardised format making assimilation challenging. Currently PyOpenWorm acts as a quick data access layer for researchers to query C. elegans anatomy and physiology. It currently incorporates data from the primary literature, WormAtlas and WormBase. This project would aim to expand the data available for model validation and for query in PyOpenWorm. Expanding the data available would aid hypothesis generation in the C.elegans research community, by allowing fast access to past observations and by facilitating the generation of more robust models through model validation tests such as SciUnit. However, ultimately it could also be a poster child to inspire an increase in standardisation and statistical accountability in the publication of biological research as a whole. As an additional project, we will start work to create "Standard Worm" which will enable researchers to compare variability in experimental worms to a standard reference space.</p>

The ‘Open Source Community Sustainability’ project is about simulating open-source community interactions to encourage conditions leading to sustainable practice. I propose to do this by creating a framework with an agent-based model powered by Large Language Models (LLMs). This project has been worked on in OREL for the past 2 years, and a variety of models and tools have been created towards it, utilizing reinforcement learning, active inference and attraction-repulsion based approaches to build several agent-based models. Thus, this framework will represent open-source communities as a system composed of interconnected components (agents), mapping the intricacies of community engagement and collaboration as well as focusing on emergent properties and overall system behaviour. This project is important to the open-source ecosystem because as the demand for open-source software continues to grow, so do the challenges associated with community management, collaboration, and sustainability. Thus, providing a framework to understand how individuals collectively contribute to community growth and success can help derive insights into effective community management, collaboration strategies, and sustainability factors.

<p>Brian is a free, open source simulator for spiking neural networks. It is written in the Python programming language and is available on multiple platforms. Brian is designed to be easy to learn and use, highly flexible and easily extensible. It is a simulator designed to save processing time and to ease the work of scientists.</p> <p>With this increased acceptability of NeuroML as a standard model for describing detailed models of neural systems, a need to implement support for NeuroML into Brian simulator has increased. With so many models now using NeuroML for describing their neural systems, there has been a demand to integrate NeuroML morphology extraction inside Brian. NeuroML also contain LEMS definitions of the ion channels and their distribution which will be extracted and used.</p> <p>This project is mainly divided into 3 parts:</p> <ul> <li><strong>Implement support to import NeuroML morphologies inside Brian 2</strong> </li> <li><strong>Add support to conveniently access to other information stored in the NeuroML file, like ion channels and their distribution information.</strong> </li> <li><strong>Test and evaluate differences between simulations of NeuroML models in Brian 2 and other simulators (such as NEURON)</strong></li> </ul>

To start with OWASP Maryam, it is a modular open-source framework based on OSINT (Open-source intelligence) designed specifically for robust data gathering from search engines and open sources efficiently and quickly. Iris, which has been recently added to Maryam and currently in beta phase, is a module with basic Natural Language Processing for clustering, sentiment analysis along with a simple meta search engine.

<p>Neural networks are a powerful tool in machine learning. An integral part of any network is its architecture. However, structuring and comparing new architectures is non-trivial. We describe a toolkit for structuring neural networks that leverages a special class of Naperian functors and functions between them. The explored class admits a canonical traversable structure which allows common neural network constructs, and shows the literal correspondence between higher-order functions and some architectures. We also show our toolkit is compatible with monadic computations necessary for training and evaluation. Our methods allow for expressive, unified, and elegant constructions.</p>