Two days. Single track. All times in IST.

..

Our Sun, while sustaining life on Earth, also constantly erupts streams of charged particles and bursts of plasma that fill the solar system and shape the space around Earth. These high energy events, known as space weather, can damage satellites, disturb power grids, and interrupt navigation and communication systems. However, forecasting these events accurately and in a timely manner remains a significant challenge. While simple empirical models are fast, they offer limited insight. Conversely, detailed physics-based simulations are accurate but too slow for real-time forecasting. Artificial intelligence is breaking this deadlock by integrating both approaches into an explainable learning pipeline.

In this talk, I will demonstrate how deep learning is advancing the forecasting of space weather drivers, particularly solar wind. By incorporating system information into the choice of architecture and training curriculum, these models provide significant enhancements in accuracy and generalizability. Additionally, feature attribution methods help clarify/discover which factors truly drive an event. I will also showcase how physics-informed neural operators can effectively emulate MHD solutions, accelerating heavy simulations from hours to seconds while maintaining physical fidelity.

Machine learning is reshaping earth system science, but its scientific value hinges on interpretability rather than prediction alone. This talk surveys our group's work on explainable AIML across the physical sciences. First, using an isotope-informed ML framework at a forested oasis on the Thar desert margin, we identify Mt. Abu as a moisture-recycling hotspot (~36% recycled vapour) and recover an unexpected non-linear control—elevated tropospheric ozone suppressing transpiration—directly from meteorological inputs. Second, combining LightGBM with SHAP attribution, we explain why summer-monsoon rainfall has risen over the eastern Indo-Gangetic Plain against the all-India decline, tracing it to irrigation-driven surface cooling that reorganises moisture convergence. Finally, turning explainability on the models themselves, we apply Random Matrix Theory to transformer weight matrices, probing how information is encoded during training and revisiting the heavy-tail versus bulk-bleeding debate over how large language models actually learn.

South Asia faces substantial challenges related to air pollution, extreme weather, and climate change. However, systematic and long-term observations remain scarce, while conventional models exhibit considerable biases and require large computational resources. Our group has been applying the AI/ML/DL approaches to bias-correct the existing model products and to develop emulators for atmospheric trace gases, aerosols, and meteorological variables. Our DL emulators show promising performance for meteorological parameters and surface ozone, although further improvements are needed in the aerosol fields. In addition, AI/ML models trained using field observations substantially outperformed several widely used global models over the Himalayan region. These findings demonstrate the strong potential of AI/ML for enhancing weather, climate, and air-quality assessments over South Asia.

Artificial Intelligence (AI) is rapidly transforming modern astronomy by enabling the efficient analysis, interpretation, and management of the enormous volumes of data generated by contemporary observational facilities. From ground-based telescopes and space missions to time-domain surveys and multi-wavelength observatories, astronomical datasets have reached a scale and complexity that often exceed the capabilities of traditional analysis techniques.

Autonomous agents increasingly plan, invoke tools, and act across enterprise systems. Two questions determine whether such agents can be trusted in production: whether an agent performs its task correctly, and whether it is authorized to perform the actions it attempts. Conventional usage telemetry — token counts, request volumes, and latency — addresses neither. This talk contends that reliability and governance are properties of the infrastructure in which agents execute, and are best enforced at that layer rather than in application middleware.

The position is examined through two independent open frameworks the author has built. The first, Synapse, is an agent runtime in which observability and evaluation are execution primitives. Owning execution, tool dispatch, and state, the runtime records every input, output, reasoning step, and tool invocation without external instrumentation, and applies online signal classifiers — for hallucination, task failure, user frustration, reduced effort, context loss, and jailbreak attempts — to live traffic rather than to a static evaluation set. Offline benchmarks characterize a system as previously configured; production traffic characterizes it as presently deployed. Sticky-variant A/B experimentation links a configuration change to its measured effect on signal frequency.

The second framework, an MCP Gateway for agent governance and provisioning, is architecturally distinct from Synapse. It fronts downstream tool servers as an OAuth 2.1 authorization server, federates human identity to an enterprise identity provider, and issues short-lived, scoped tokens. Each tool call traverses an authorization pipeline of identity classification, policy evaluation, and human-in-the-loop review. A composite-identity model ensures that an agent never holds a downstream credential; the gateway injects the appropriate credential at egress. Autonomous agents operate without a live human under an owner's vaulted, revocable grant.

The two systems are independent, not a single pipeline: each instantiates the thesis that agent reliability and safety belong, respectively, in the runtime and the access layer. The talk presents both architectures, their principal design trade-offs, and the lessons drawn from implementation.

Autonomous coding agents are increasingly embedded in software engineering workflows, yet enterprises still lack reliable ways to determine whether the tokens, context, memory, and tool calls consumed by these agents produce valid outcomes. Current agent telemetry often measures usage: requests, token volume, cache reads, cache writes, and tool activity. These signals are useful, but they do not answer the more important question: did the agent understand the user's intent, plan effectively, use the right context, invoke the right tools, generate correct code, verify its work, and avoid downstream rework?

This talk proposes Agent Efficacy Management, a closed-loop framework for evaluating and improving multi-step autonomous agent architectures using process intelligence. The approach treats an agent-assisted task as an event log that connects user prompts, agent plans, memory and cache behavior, file reads, tool calls, code edits, tests, corrective prompts, Git activity, and review outcomes. From these traces, teams can measure token efficacy: validated outcome quality per token-adjusted effort, including human rework and correction loops.

The proposed framework introduces practical measures such as intent adherence, first-time-right behavior, context precision, cache yield, verification discipline, rework token ratio, and outcome durability. These metrics directly support the study of planning, memory, and tool-augmented reasoning by showing when an agent's reasoning process leads to a reliable result and when it creates avoidable waste or failure.

Low-efficacy patterns can then be converted into interventions such as better prompting guidance, repository-specific rules, reusable agent skills, verifier subagents, and human-approved governance workflows. The contribution is a process-oriented method for making autonomous coding agents more accountable, resilient, and improvable in real software environments. By connecting agent execution traces to outcome quality and repeatable improvement actions, Agent Efficacy Management provides a bridge between autonomous agent systems, software engineering governance, and enterprise process intelligence.

The rapid growth of glacial lakes in the last decade, especially in the Hindu Kush Himalayan region is a source of grave concern as the larger ones often spawn outburst Glacial Lake Outburst Floods (GLOFs). Since GLOFs are known to cause widespread damage to all in their path, understanding and modelling the factors causing them may provide solutions to averting such disasters. This talk will focus on some machine-learning based approaches for studying the various aspects towards understanding such events based on multimodal data acquired primarily through remote sensing and complemented with reanalysis data. The applications range from accurate determination of glacial lakes from satellite images, tracking their growth in terms of area, modelling climatic factors identified as key contributors of GLOFs and finally, prediction of the flood pathway. Preliminaries would be covered at the beginning while the talk will conclude by discussing some ongoing work and intended research directions.

This talk introduces the emerging role of machine learning (ML) in advancing ocean science through data-driven modeling. It highlights how ML techniques can complement traditional numerical models by integrating diverse observational and model-generated datasets to improve the representation of oceanic processes, reduce model uncertainties, and enhance predictive capabilities in oceanography and climate science.

The talk presents two case studies demonstrating the integration of observations, numerical model outputs, and reanalysis products to develop improved ocean state estimates.

The first case study focuses on reconstructing surface ocean partial pressure of carbon dioxide (pCO2) in the Indian Ocean. Although high-resolution ocean biogeochemical models provide a valuable framework for investigating carbon cycle variability, their pCO2 simulations often exhibit biases arising from uncertainties in parameterizations, physical processes, and surface and boundary forcing. These biases propagate into estimates of air-sea CO2 fluxes, limiting our understanding of regional carbon uptake. To address this challenge, we developed a hybrid high-resolution pCO2 data product by combining observations from the Surface Ocean CO2 Atlas (SOCAT), Indian scientific cruises, a regional ocean biogeochemical model, and an advanced machine-learning framework. Validation against independent observations, including the BOBOA mooring (90°E, 15°N), gridded SOCAT products, and other ML-based reconstructions, demonstrates an approximately 40 ± 3.31% reduction in root mean squared error relative to the original model. This hybrid product significantly improves the representation of observed pCO2 variability and reduces uncertainties in regional air-sea CO2 flux estimates.

The second case study presents INCOIS_TA, the first high-resolution (1/12°), gap-free, monthly surface total alkalinity (TA) data product for the North Indian Ocean spanning 1993–2020. Developed using an ensemble machine-learning approach trained on ship-based observations and reanalysis variables, including sea surface temperature, salinity, and mixed-layer depth, the data product captures regional and temporal variability in ocean buffering capacity. By providing both TA estimates and uncertainty information, INCOIS_TA offers a valuable benchmark for model validation, bias correction, carbonate system studies, and improved assessments of ocean acidification, carbon cycling, and climate feedbacks.

The exponential growth of astronomical data from various surveys has made Artificial Intelligence (AI) a necessary tool for extracting and analyzing data. Deep Learning (DL) is becoming increasingly appealing compared to traditional Machine Learning (ML) because it can learn features from data without significant manual feature engineering. A major challenge in this observational science is the discrepancy between theoretical models (simulations) and observed data. The outputs estimated using DL models trained on simulated data often fail to generalize well to real observations due to the distributional shift of learned features across different data sources. This distributional shift causes significant performance degradation when models trained on one dataset are tested on a different one. A further limitation of DL relative to ML is its black-box nature, which makes it difficult to interpret why a model makes a given prediction. In this talk, we explore various domain adaptation methods to address distributional shifts and apply explainability techniques, specifically saliency maps, to examine how learned representations and feature usage change before and after adaptation.

PRL’s Indian Lidar Network (ILIN), a nationwide network with 17 Lidar observatories covering different regions of India that work for the investigation of the vertical structure and dynamics of atmospheric clouds, boundary layer, and related atmospheric processes across the Indian region. The ILIN aims to advance atmospheric science by examining boundary layer evolution, low- and mid-level cloud formation, and their dynamical interactions. The network also aims to provide essential ground-based observations for the validation of satellite-derived cloud and Fog products and AI/ML algorithms, along with ground-truth observational input for regional weather and climate models.

As structured light fields traverse through the multimode fiber (MMF), the inherent randomization of spatial phases is driven by intermodal coupling, which transforms the incident wavefront into a stochastic speckle distribution. While this process appears to scramble the initial state, the resulting intensity pattern remains a deterministic signature of the input mode's spatial properties. In this work, we demonstrate a robust framework for identifying constant-size-scaled 64 distinct Laguerre-Gaussian (LG) beams, after transmission through a 1 m MMF. By utilizing these 64 distinct modes whose ℓ values ranging from −32 to +32 (excluding zero), we achieve uniform fiber excitation across all the modes. To decode the high-order scattering at the fiber output, a modified AlexNet-based convolutional neural network (CNN) is employed to classify the emergent speckle distributions. The model achieves 99.84% classification accuracy, demonstrating that deep learning can effectively map the nonlinear transformations of spatial modes within a complex medium. To validate the practical utility of this mapping, we performed pixel-level image transmission and recovery, where the retrieved images exhibited high structural fidelity. By utilizing 64 distinct spatial modes instead of binary states of 0 and 1, this approach increases information density and ensures that the information can be reliably recovered even after being scrambled by a complex MMF channel. This approach highlights the potential of CNN-enhanced optical systems for information recovery and secure pattern recognition for optical communication.

Machine learning (ML) has transformed our approach to modeling complex physical systems by not only enabling faster and more accurate simulations but also creating entirely new computational pathways for the atomic-scale discovery of compounds with distinct properties. Unlike many other domains, atomistic ML benefits from rich physical laws. For instance, atomic structures transform predictably under rotations, translations, and inversions, and these symmetries can be embedded into models. Such physics-informed approaches have already advanced the prediction of critical material properties and help bypass the steep cost of ab initio calculations for quantities ranging from potential energy surfaces to dipole moments and polarizability tensors.

In this talk, I will present several approaches for incorporating physical domain knowledge into machine learning, either through physics-aware input representations or physically motivated model architectures, thereby making black-box ML models more interpretable. I will then show how these approaches can be used not only to predict material properties, but also to uncover the mechanisms connecting atomic structure to observable physical observables.

Although the examples will focus on atomistic systems, the underlying ideas are broadly applicable to machine learning problems involving point cloud data.

TBD.

While continuous diffusion models have achieved remarkable success, discrete diffusion offers a unified framework for jointly modeling text and images. Beyond unification, discrete diffusion provides faster inference, finer control, and principled training-free guidance, making it well-suited for posterior sampling. Existing approaches to posterior sampling using discrete diffusion face severe challenges: derivative-free guidance yields sparse signals, continuous relaxations limit applicability, and split Gibbs samplers suffer from the curse of dimensionality. To overcome these limitations, we introduce Anchored Posterior Sampling (APS) [ICML 2026], built on two key innovations: quantized expectation for gradient-like guidance in discrete embedding space, and anchored remasking for adaptive decoding. APS achieves state-of-the-art performance among discrete diffusion samplers on both linear and nonlinear inverse problems across the standard image benchmarks. We demonstrate the generality of APS through training-free stylization and text-guided editing. We further apply APS to a large-scale diffusion language model, showing consistent improvement in question answering. Project page: https://anchored-discrete-ps.github.io/

With the advent of large volumes of data, heliophysics has been leading with the presence of high cadence, high resolution and long term data. From the solar interior to the near-Earth plasma interactions, machine learning algorithms have been developed for applications across the board, resulting in the generation of a wealth of understanding about this Sun. In this talk, I shall present several applications of machine learning in heliophysics, and how they have advanced our understanding of various phenomena on the Sun. I shall conclude the talk by presenting the solar physics foundation model Surya, and zero-shot results from our science validation exercise.

TBD.

Government service rules and administrative policies in India are distributed across thousands of Office Memoranda, circulars, and gazette notifications issued by the Department of Personnel and Training (DoPT), covering areas such as leave, pension, allowances, deputation, promotions, and performance incentives. Retrieving and interpreting the appropriate rule for a specific service matter is often time-consuming and requires considerable domain knowledge, while conventional keyword-based search systems typically return ranked lists of documents rather than precise, context-aware answers.

RuleEasy, a multilingual Retrieval-Augmented Generation (RAG) system designed to provide accurate, source-grounded interpretation of government service rules through a conversational interface. The system is built on WARC-GPT, an open-source framework developed by the Harvard Library Innovation Lab, and indexes text chunks extracted from DoPT policy documents using the e5-large-v2 embedding model and ChromaDB. Responses are generated by Llama 3.1 (70B), deployed locally through Ollama on the PARAM VIKRAM 1000 high-performance computing facility at the Centre for National Information Technology (CNIT), Physical Research Laboratory (PRL).

To overcome cross-lingual retrieval limitations associated with a monolingual embedding model, a lightweight Python-based query translation module was developed, achieving approximately 90% retrieval success across all ten DoPT policy domains. By integrating semantic retrieval, multilingual query processing, conversational interaction, and verifiable source attribution, RuleEasy improves access to complex government service rules and demonstrates the practical application of open-source RAG technologies for multilingual knowledge retrieval in the Government sector.

Technological advancement often relies on the development of advanced materials. Therefore, the discovery of new materials with desired properties is an important but challenging task as trial-and-error methods of synthesizing novel compounds in the laboratory can be extremely labor, time, and cost-intensive. Computational methods combined with autonomous design strategy have the potential to reduce all three factors, significantly. In this talk, I will focus on the rapid discovery of a class of energy materials having low lattice thermal conductivities and promising thermoelectric performance using advanced computational methodologies. First, I will show how the rational design of an automatic workflow can accelerate the discovery of thermodynamically stable compounds based on high-throughput density functional theory calculations. Then, I will demonstrate how a machine-learning model based on the crystal graph convolutional neural network can help discover the unsuspected compounds in the same family of materials, that would otherwise be difficult to do in conventional ways. Recent experimental validation of some our predictions underpins the success and usefulness of computationally guided materials design approach. Lastly, I will briefly discuss efficient materials innovation strategy that can be achieved through the design of a closed-loop autonomous infrastructure capable of performing discovery, synthesis, and characterization of novel materials.

References: (1) Pal K, Xia Y, Shen J, He J, Luo Y, Kanatzidis MG, Wolverton C. Accelerated discovery of a large family of quaternary chalcogenides with very low lattice thermal conductivity. npj Computational Materials. 2021 May 10;7(1):82. (2) Pal K, Park CW, Xia Y, Shen J, Wolverton C. Scale-invariant machine-learning model accelerates the discovery of quaternary chalcogenides with ultralow lattice thermal conductivity. npj Computational Materials. 2022 Mar 15;8(1):48. (3) Singh N, Ubaid M, Nayak PK, He J, Ghosh D, Wolverton C, Pal K. Data-driven discovery of novel high-performance quaternary chalcogenide photovoltaics. Physical Review Applied. 2026 Mar 20;25(3):034073.

Deep learning is increasingly used to study complex data from particle colliders, but it is often difficult to understand what these models actually learn. In this talk, I will discuss how ideas from physics can be built into machine learning models to make them more stable and interpretable. Focusing on jets produced in high-energy collisions, I will show how such physics-aware models can help connect AI-based predictions with physically meaningful patterns.

Recent planetary missions have produced an unprecedented volume of high-resolution imaging, spectral, radar, and topographic data, transforming our understanding of the Moon and Mars. This talk presents the perspective of a planetary scientist and highlights the major unsolved scientific questions in lunar and Martian exploration where AI and ML have the potential to fundamentally change how science is conducted. Using examples from current research on the Moon and Mars, the talk will discuss opportunities for AI-driven approaches in planetary sciences. It will also outline key challenges that require further discussions. The presentation will focus on bridging the gap between planetary science and AI/ML by identifying scientifically meaningful problems where AI/ML techniques can accelerate discovery and support future robotic and human exploration of the Solar System.

Star clusters are the fundamental building blocks of galaxies and serve as natural laboratories for studying stellar evolution, Galactic structure, and the assembly history of the Milky Way. Over time, clusters dissolve through internal dynamical evolution and external tidal interactions, dispersing their members into the Galactic field. Identifying these dispersed siblings is essential for accurately determining cluster properties and reconstructing their evolutionary histories. The advent of large-scale surveys such as Gaia, combined with modern machine learning techniques, has opened new avenues for addressing these challenges by enabling robust cluster membership determination and the recovery of distant stellar siblings. In this talk, I will highlight how AI/ML methods are transforming our understanding of star clusters and their long-term evolution, and present case studies demonstrating their applications to cluster membership, stellar stream identification, and the reconstruction of dispersed stellar populations.

Several ground- and space-based solar facilities have become operational in the last two decades, all of which generate an enormous volume of multi-dimensional datasets that require novel and efficient tools to analyze them in a timely manner. In this talk, I shall describe the concept of k-means and DBSCAN, which are popular unsupervised clustering algorithms widely used in the domain of machine learning, and their relevance to the investigation of small-scale inhomogeneities in sunspots.

Sun is the primary source of energetics that drive vibrant phenomena in the terrestrial space environment. Multitude of processes taking place from the Sun, interplanetary medium, magnetosphere and atmosphere, constitutes “Space Weather”. The continuous steam of radiation, particles and transient explosive events from Sun are the ‘source’ of space weather, consequent modifications in the near earth space environment comprises its ‘effects’. These changes severely ‘impact’ the space and ground based technologies and infrastructure such as, communication, aviation, navigation, satellite hardware and operations, human in space, power grids, mining among others. Since the space weather is a system of systems involving multitude of phenomena, which are interlinked among each other, its modelling demands suitable approaches to quantify and reproduce complex processes. Numerous efforts have been in place for several decades through data driven, empirical and theoretical approaches. However, there exists several limitations to suitably simulate and quantify various physical phenomena for efficient prediction. The AI/ML is a powerful tool to fill this gap, to analyse large datasets, address the complex processes in order to arrive at appropriate methodologies to meet the requirements of day to day technologies. Recent results on the ‘cause’, ‘effect’ and ‘impact’ of space weather using AI/ML approaches will be presented. The efficacy of AI/ML to bring transformational changes in the space weather prediction capabilities will be discussed.

Weather prediction has traditionally relied on physics-based numerical models, which remain fundamental to operational forecasting. Recent advances in artificial intelligence, driven by the availability of large Earth observation datasets and increased computational capability, have transformed the way atmospheric processes are modeled and predicted. This talk presents an overview of the evolution of AI architectures for weather prediction, from conventional machine learning and deep learning models to transformers, foundation models, and emerging physics-informed approaches. The discussion highlights the motivation behind these developments, their strengths and limitations, and their impact on modern weather forecasting. The talk concludes with current challenges and future directions toward developing accurate, reliable, and physically consistent AI systems for Earth system science.

The Lunar surface has been bombarded with several million impacts since the last 4.5 billion years, which form millions of craters. Those craters' diameters range widely from a few meters to a few 1000s of kilometres. The number of recent impacts occurred on the Moon and their location is not quantified till date. Even though the lunar surface is mapped by multiple high-resolution instruments on satellites like Lunar Reconnaissance Orbiter, Chandrayaan-1 & 2, with spatial resolution as fine as 0.5 meter/pixel to 0.25 meter/pixel, this quantification remains unsolved. A unique way to identify the recent or fresh impacts is through their ejecta material, which splashed isotropically and radially away from the impact site. Such fresh craters ejecta displays very high albedo with bright, spoke-like rays, composed of streaks of pulverized rock and dust radiating from the impact point, because they haven’t yet been darkened by space weathering. In this study, we are mapping such fresh craters on the Moon, globally, using deep learning. This research presents an automated pipeline to detect and map fresh craters between 40°N and 40°S using Lunar Reconnaissance Orbiter Narrow Angle Camera (LRO NAC) imagery. A YOLOv11n segmentation model was trained on 876 images at 1024x1024 resolution, each showing a clear example of ejecta morphology. This segmentation was used to map the irregular and radial areas where the masks capture the ejecta pattern’s shape and extent. The code developed includes many unique features, such as deep learning for fresh ejecta of various sizes, learning craters under varying illumination geometries, and multiple iterations leading to the final detection of fresh craters. In total, our study region contains 4,00,000 images, each with a size of ~500 MB. Processing such large datasets requires high computational demands which was handled by the PRL’s Param Vikram 1000. The PRL HPC houses 108 compute nodes, 7,296 CPU cores, 2,76,480 GPU cores, 74 TB of RAM and 1 PB of storage. Images are downloaded via the ODE REST API, georeferenced using GDAL tools and then passed to GPU-parallel inference. The results are output as georeferenced images and ejecta blanket shape files. Our preliminary test run on 1,25,000 LRO-NAC images detected nearly 11,00,000 craters. The code is undergoing extensive deep learning through feedback from earlier detection results. In this talk, we will show the preliminary results from the 0°-1° (latitude and longitude) region of fresh craters detected by our PRL-M_FC model.

Recent advances in foundation models have transformed fields such as Earth observation and medical imaging, but their application to planetary science remains largely unexplored. This talk presents two recent contributions to machine learning for planetary science: Mars-Bench, a benchmark for evaluating foundation models on Mars science tasks, and MOMO, a multi-sensor foundation model for Mars orbital data. Together, these works provide a framework for developing and assessing foundation models for Mars exploration and scientific discovery.

High-stakes environments are often riddled with ambiguity, vagueness, and logical gaps that fundamentally challenge traditional artificial intelligence. This methodology mitigates these risks via a dual approach: utilizing theorem proving to formalize complex logic, alongside specialized verification tracks to certify neural reasoning. This talk demonstrates a compiler-driven architecture that transforms intricate operational specifications into provably correct execution, ensuring absolute reliability for mission-critical applications.

While large language models (LLMs) excel at natural language generation, they struggle with complex reasoning and multistep planning tasks in high-stakes environments. To address this limitation, we present a neurosymbolic framework integrating domain-adapted LLMs with symbolic knowledge graphs, graph-based reasoners, and constraint-aware planning modules. Aligning with the demand for resilient AI systems, this architecture advances Knowledge-Grounded Planning and Reasoning by utilising LLMs for initial plan formulation while strictly grounding subsequent refinement in structured domain knowledge. This integration enforces safety standards, ensures regulatory compliance, and maintains logical consistency. By combining generative flexibility with formal logic, this hybrid approach bridges the gap between unstructured language and symbolic reasoning, offering a scalable, trustworthy foundation for complex, constraint-driven decision support.

..