Predict shell formation, stability, magnetic manipulability, and triggered release where classical regime maps and exhaustive CFD fall short.
Observable-aware AI for liquid-liquid encapsulation and release.
INTERACT-Capsules turns high-speed videos, multiphase simulations, and physics-aware operator learning into an experimental design tool for four-fluid encapsulation systems.
Fuse simulation, video-derived interface trajectories, neural operators, and active learning into a calibrated design assistant.
A lab-facing tool that recommends operating conditions, explains failure modes, and selects the next most informative experiments.
Why it matters
Encapsulation is powerful, but still too heuristic.
Liquid-liquid encapsulation can protect fragile or responsive cargo without solidifying the shell or exposing payloads to harsh processing. That makes it attractive for drug delivery, tissue engineering, soft robotics, microscale synthesis, and responsive material systems.
The design challenge is that small changes in impact height, layer thickness, viscosity ratio, density contrast, shell composition, magnetic loading, or cargo structure can flip outcomes from stable wrapping to trapping, penetration, rupture, asymmetry, or release only under impractical actuation.
Computational stack
Four layers from raw observables to prospective design.
Simulation backbone
Adaptive multiphase CFD provides structured training data and counterfactual cases across impact velocity, shell thickness, viscosity, density, interfacial tensions, and magnetic descriptors.
Video state inference
High-speed videos become interface trajectories: lamella formation, penetration depth, neck radius, shell closure, eccentricity, shell thickness, trapping signatures, and release dynamics.
Multimodal prediction
Interface-aware neural operators combine fluid properties, controls, and early observables to estimate regime class, success probability, geometry, retention, uncertainty, and actuation thresholds.
Inverse design
Bayesian optimization and information-aware acquisition recommend operating windows for target capsule behaviors while surfacing where the model is confident, uncertain, or extrapolating.
Experimental scope
Three families, increasing in difficulty.
Baseline and constrained-layer encapsulation
Single-shell wrapping and constrained interfacial layers for learning the core relations among impact energy, shell thickness, viscosity ratio, and encapsulation success.
Multilayer and magnetoresponsive shells
Double- and triple-layer liquid cargos, ferrofluid-containing shells, and downstream behavior that depends on layered morphology and interfacial integrity.
Liquid-wrapped hydrogels and triggered release
Magnetically responsive and non-magnetic hydrogel cargos with outputs for transport stability, selective release, and manipulation thresholds.
Success criteria
Clear validation gates for a tool experimentalists can trust.
Workplan
An 18-month route to a hardened lab-facing release.
Standardize data schemas, assemble prior datasets, stand up simulations, and implement segmentation and contour extraction.
Train the first Family A multimodal surrogate and release an internal prototype for baseline and constrained-layer planning.
Improve simulation-to-experiment transfer, uncertainty calibration, and the first prospective recommendation engine.
Extend to multilayer and ferrofluid-containing shells, integrate magnetic descriptors, and benchmark cross-family transfer.
Add hydrogel release planning and compare model-guided campaigns against heuristic exploration.
Harden the software, finish prospective validation, prepare documentation, model cards, datasets, and manuscripts.
Risk management
Staged scope when harder regimes resist.
Simulation-to-experiment mismatch
Simulations act as structured priors rather than ground truth, with calibration on experiments and explicit uncertainty for extrapolation.
Complex multilayer and magnetic release physics
Family A alone yields a useful tool, while Families B and C extend scope as data and representations mature.
Rare but important failure modes
Active learning and disagreement-driven acquisition target under-sampled failures instead of relying on passive data collection.
Team
Computational AI-for-science meets experimental interfacial matter.
The project pairs Kevin Dsouza's software, machine-learning, and multiscale modeling background with Dr. Sirshendu Misra's experimental platform for liquid-liquid encapsulation, ferrofluid-wrapped droplets, and magnetic release.
KD
Kevin Dsouza
Primary applicant; AI-for-science software and modeling lead
Kevin brings experience in AI-native research software, multiscale modeling, protein design workflows, climate and landscape optimization, and biological representation learning. His role in INTERACT-Capsules is to lead the computational stack, uncertainty-aware models, active design engine, and experimentalist-facing software.
SM
Dr. Sirshendu Misra
Equal co-applicant; experimental interfacial platform lead
Dr. Misra is a Ramanujan Fellow Faculty member at the Indian Institute of Science, Bengaluru, with a research program spanning wetting, thin films, surface science, soft matter, bio-fluid dynamics, and liquid-liquid encapsulation. His group anchors the high-speed imaging, experiment design, and prospective validation.
Selected papers
Project-relevant foundation and team track record.
Liquid-liquid encapsulation: Penetration vs. trapping at a liquid interfacial layer
Misra, Banerjee, Mitra · ACS Applied Materials & Interfaces · 2023
A versatile multilayer liquid-liquid encapsulation technique
Banerjee, Misra, Mitra · Journal of Colloid and Interface Science · 2025
Manipulation of ferrofluid-wrapped drops: translation, coalescence, and release
Gunjan, Banerjee, Misra, Mitra · Advanced Materials Interfaces · 2023
Magnetic manipulation of liquid-wrapped hydrogels
Banerjee, Misra, Mitra · Soft Matter · 2025
Learning representations of chromatin contacts using a recurrent neural network
Dsouza et al. · Nature Communications · 2022
Latent representation of the human pan-celltype epigenome
Dsouza, Li, Bhargava, Libbrecht · IEEE/ACM TCBB · 2021
BoreaRL: A Multi-Objective Reinforcement Learning Environment for Climate-Adaptive Boreal Forest Management
Dsouza et al. · arXiv · 2025
Structuring Collective Action with LLM-Guided Evolution
Dsouza et al. · arXiv · 2025
Talks and public presence
Relevant scientific communication and online footprint.
Representation learning for biology
Kevin's invited and conference talk thread across 4DN, RECOMB, MLCB, epigenetics meetings, and online venues.
Talks archiveLiquid encapsulation public coverage
Early public coverage of the impact-driven liquid encapsulation platform and its potential for rapid, low-energy drug-delivery workflows.
Read coverageOpen implementation
The project codebase includes data schemas, validation scripts, simulation planning, multimodal training, uncertainty calibration, recommendation tooling, and MVP documentation.
Open GitHub repo