Portfolio item number 1
Published:
Short description of portfolio item number 1
Posts by Collection
portfolio
Portfolio item number 2
Published:
Short description of portfolio item number 2
publications
talks
Fully Decoupled Time-Marching Schemes for Incompressible Fluid/Thin-Walled Structure Interaction
Published:
Conference presentation on fully decoupled time-marching schemes for incompressible fluid and thin-walled structure interaction. The schemes combine a projection method in the fluid with a Robin-Neumann treatment of the interface coupling, so that the fluid velocity, the fluid pressure, and the structure displacement are computed sequentially, each only once per time step. The talk discussed energy-based stability properties of the resulting schemes with respect to the added-mass effect and illustrated their accuracy and efficiency on numerical benchmarks.
Recent Developments in Explicit Robin-Neumann Schemes for Fluid-Structure Interaction
Published:
Conference presentation on recent developments in explicit Robin-Neumann schemes for fluid-structure interaction. These schemes replace the classical Dirichlet-Neumann coupling with a Robin condition on the fluid side that accounts for the inertia of the structure, so the fluid and the structure are solved only once per time step without sub-iterations. The talk reviewed their stability and accuracy with respect to the added-mass effect and presented extensions such as fully decoupled velocity-pressure-structure splittings.
Splitting Schemes for Fluid/Thin-Walled Structure Interaction Using Unfitted Meshes
Published:
Poster presentation on splitting schemes for fluid and thin-walled structure interaction using unfitted meshes. The work extends explicit Robin-Neumann coupling to configurations where the structure is not matched by the fluid mesh, with the interface conditions enforced weakly through Nitsche’s method. This combines the efficiency of loosely coupled time-marching with the flexibility of unfitted discretizations, which avoid remeshing or deforming the fluid mesh as the structure moves.
Nitsche-XFEM for the Coupling of an Incompressible Fluid with Immersed Thin-Walled Structures
Published:
Conference presentation on Nitsche-XFEM methods for coupling an incompressible fluid with immersed thin-walled structures. The fluid is discretized on a mesh that is not fitted to the structure, with extended finite elements allowing the velocity and pressure to be discontinuous across the immersed interface and Nitsche’s method enforcing the fluid-solid coupling on cut elements. Several coupling schemes with different degrees of fluid-solid time splitting, from implicit to explicit, were compared on numerical examples with static and moving interfaces.
Nitsche-XFEM for the Coupling of an Incompressible Fluid with Immersed Thin-Walled Structures
Published:
Seminar talk on Nitsche-XFEM methods for coupling an incompressible fluid with immersed thin-walled structures. The talk presented an unfitted-mesh approach in which a Lagrangian thin-walled structure cuts through an Eulerian fluid mesh, using extended finite elements to capture discontinuities in the fluid velocity and pressure and Nitsche’s method to impose the interface conditions. It also discussed stabilization for arbitrary interface intersections and the trade-offs between implicit, semi-implicit, and explicit time-splitting of the fluid-solid coupling.
Nitsche-XFEM Formulations and Splitting Schemes for the Coupling of an Incompressible Fluid with Immersed Thin-Walled Structures
Published:
Conference presentation on Nitsche-XFEM formulations and splitting schemes for coupling an incompressible fluid with immersed thin-walled structures. The talk combined an unfitted-mesh spatial discretization, in which extended finite elements capture fluid discontinuities across the structure and Nitsche’s method imposes the interface coupling, with time-splitting schemes that solve the fluid and the structure sequentially. The stability and accuracy of the resulting semi-implicit and explicit schemes were discussed and illustrated with numerical examples.
Coupling Schemes for the FSI Forward Prediction Challenge: Comparative Study and Validation
Published:
Conference presentation on coupling schemes for the FSI Forward Prediction Challenge, with a comparative study and validation focus. The challenge is built around an experimental fluid-structure interaction benchmark, providing measured data against which computational predictions can be assessed. The talk compared implicit, semi-implicit, and explicit coupling schemes in terms of accuracy and computational cost and validated the simulations against the experimental measurements.
Coupling Schemes and Unfitted Mesh Methods for Fluid-Structure Interaction
Published:
Conference presentation on coupling schemes and unfitted-mesh methods for fluid-structure interaction, summarizing the main results of the PhD thesis completed at Université Pierre et Marie Curie and Inria. The talk covered explicit Robin-Neumann splitting schemes for thin-walled structures and Nitsche-XFEM discretizations for structures immersed in an incompressible fluid. Presented as a participant in the ECCOMAS PhD Olympiad 2017.
Numerical Approximation of the Electromechanical Coupling in the Left Ventricle with Inclusion of the Purkinje Network
Published:
Workshop talk on numerical approximation of electromechanical coupling in the left ventricle, including the Purkinje network. The model couples electrical propagation along a one-dimensional Purkinje network with the three-dimensional electrophysiology of the ventricular muscle, which in turn drives the active mechanical contraction of the tissue. The talk presented a segregated numerical strategy for this multiphysics problem and discussed how including the Purkinje network affects the simulated activation and contraction of the ventricle.
Improving Exploration in Policy Gradient Search
Published:
Workshop talk on improving exploration in policy-gradient search for symbolic optimization problems. When neural networks trained with reinforcement learning are used to search combinatorial spaces of mathematical expressions, the search can suffer from early commitment and initialization bias, both of which limit exploration. The talk introduced two remedies, a hierarchical entropy regularizer and a soft length prior, and showed that they improve performance and sample efficiency and yield simpler solutions on symbolic regression tasks.
Discovering Symbolic Policies with Deep Reinforcement Learning
Published:
Spotlight presentation on discovering interpretable symbolic policies with deep reinforcement learning. The approach, deep symbolic policy, uses an autoregressive recurrent neural network trained with a risk-seeking policy gradient to generate control policies written as short mathematical expressions, and scales to multi-dimensional action spaces through an “anchoring” procedure that distills pre-trained neural policies one action dimension at a time. Across eight benchmark control environments, the discovered symbolic policies outperformed seven state-of-the-art deep reinforcement learning algorithms in average rank and normalized reward despite their dramatically reduced complexity.
Deep Symbolic Optimization: A Framework for Symbolic Optimization Using Deep Learning
Published:
Workshop talk introducing Deep Symbolic Optimization as a framework for symbolic optimization using deep learning. DSO uses a recurrent neural network to generate candidate solutions as sequences of discrete symbols, such as mathematical expressions, and trains it with a risk-seeking policy gradient that focuses learning on the best-performing samples. The talk illustrated the framework on symbolic regression and interpretable control, two tasks where compact, human-readable solutions are especially valuable.
Intracardiac Electrical Imaging Using the 12-Lead ECG: A Machine Learning Approach Using Synthetic Data
Published:
Conference presentation on intracardiac electrical imaging from 12-lead ECGs using machine learning trained with synthetic data. Because paired recordings of surface ECGs and intracardiac electrical activity are scarce, the approach trains deep neural networks on a dataset of more than 16,000 cardiac electrophysiology simulations to reconstruct activation maps and transmembrane voltages from the standard 12-lead ECG. The simulated dataset was released through the LLNL Open Data Initiative to support further research on non-invasive cardiac imaging.
Deep Symbolic Optimization for Scientific Discovery
Published:
Symposium talk on deep symbolic optimization as a framework for scientific discovery across symbolic regression, control, and other structured design problems. DSO casts the search for compact symbolic solutions, such as equations or control policies, as a sequential decision problem in which a neural network generates candidates that are rewarded by how well they explain data or solve a task. The talk surveyed the main components of the framework and highlighted results such as the unified deep symbolic regression method that won the real-world track of the 2022 SRBench symbolic regression competition.
Deep Symbolic Optimization: Reinforcement Learning for Equation Discovery
Published:
Keynote talk on deep symbolic optimization and reinforcement learning for equation discovery at AI for Science 2025. The talk presented DSO as a reinforcement learning approach to symbolic mathematics, in which a neural network learns to generate equations that are rewarded by how well they fit data. It traced the framework’s development from deep symbolic regression to broader applications such as interpretable control policies and generative design in hybrid discrete-continuous spaces.
Autoresearch Discovery of Interpretable Filter Rules for Antibody Binder Classification
Published:
Invited talk to NVIDIA’s Digital Biology / BioNeMo and Fundamental Generative AI Research teams on an “autoresearch” loop in which an LLM acts as an automated experimenter: it proposes interpretable filter rules over signals already available in a design pipeline (AlphaFold3 confidence, ProteinMPNN, antibody structural self-consistency), evaluates them under strict held-out protocols, logs each result, and iterates. Across several benchmark datasets the loop discovers compact, human-readable scoring rules that reach parity with supervised ML and state-of-the-art LLM prompting, with no model training and no LLM call at inference. The talk closes by framing the loop as symbolic regression with an LLM in the experimenter’s seat, connecting it to the interpretable-ML and equation-discovery line of work.
LLM-Guided Integer Linear Programming for Antibody Library Design
Published:
Conference talk on combining large language models with integer linear programming to design antibody libraries under constraints. In the pipeline, the LLM is scoped to emitting structured-JSON priors, and a deterministic ILP solver uses them to select a library of antibody variants that satisfies the design constraints. Keeping the optimization in the solver makes the resulting libraries auditable and easy for a domain expert to inspect or override, building on the ProtLib-Designer framework for antibody library design.
Reinforcement Learning for Antibody Sequence Infilling
Published:
Conference talk, joint work with Chak Shing Lee, Conor F. Hayes, and Denis Vashchenko, on a framework that combines a pretrained antibody infilling language model with reinforcement learning to optimize functional properties. The model generates specific antibody regions, such as CDR loops, within full sequences and is fine-tuned either online with REINFORCE and PPO under KL regularization, or offline with Direct Reward Optimization on static experimental datasets. Across design tasks including binding affinity, immunogenicity, and expression, the approach improves alignment with measured biophysical properties and outperforms likelihood-only baselines.
teaching
Teaching experience 1
Undergraduate course, University 1, Department, 2014
This is a description of a teaching experience. You can use markdown like any other post.
Teaching experience 2
Workshop, University 1, Department, 2015
This is a description of a teaching experience. You can use markdown like any other post.