Latest news with #SandeepSukumaran
Time of India
a day ago
- Science
- Time of India
Cloud computing: IIT Delhi AI model forecasts monsoon 18 days in advance
New Delhi: Two recent studies by IIT Delhi have demonstrated how artificial intelligence (AI) can enhance weather forecasting in India. In the first study, researchers used an AI model to predict monsoon patterns up to 18 days in advance — something that current models struggle to do. Tired of too many ads? go ad free now In the second study, another team compared four advanced AI weather models for cyclone tracking and found that they could predict cyclone paths with high accuracy, within 200km, up to four days ahead, and in just seconds. According to researchers, these studies show how AI can make weather forecasting faster, cheaper and more reliable, which could aid in saving lives and helping with better planning in the face of approaching big weather events. In one study, doctoral candidate KM Anirudh, working under professors Sandeep Sukumaran and Hariprasad Kodamana at IIT Delhi, has achieved major progress in monsoon prediction by applying transformer neural networks — the same type of artificial intelligence technology that powers systems like ChatGPT. The research team trained their model using 25 years of high-resolution satellite rainfall data. This allowed the system to accurately forecast the monsoon intraseasonal oscillation (MISO) — a key driver of monsoon variability that recurs every 30 to 60 days — with a lead time of up to 18 days. Current numerical weather prediction models struggle to forecast monsoon rainfall this far in advance. In contrast, the transformer-based AI model successfully captured these shifts up to two weeks ahead, while requiring significantly less computational power. In the other study, doctoral researcher Pankaj Lal Sahu, also guided by the same professors, conducted a comprehensive assessment of four leading AI-based global weather prediction models, comparing them with conventional numerical weather prediction (NWP) systems. Tired of too many ads? go ad free now The AI systems demonstrated remarkable capability in 96-hour cyclone track forecasting, maintaining positional accuracy within 200 kilometres while completing computations in seconds, rather than hours. Notably, the AI models successfully internalised complex atmospheric dynamics without explicit programming of physical equations, achieving this through advanced machine learning techniques alone. According to researchers, traditional weather models use physics and need huge supercomputers with hundreds of processors to make forecasts. But once they're trained, machine learning models can make forecasts in just seconds using only one powerful computer chip. This makes them promising for predicting dangerous weather like tropical cyclones (hurricanes and typhoons). "As extreme weather events become more frequent due to climate change, such AI-powered forecasting tools may prove indispensable for vulnerable communities worldwide," said professor Hariprasad Kodamana of IIT Delhi's chemical engineering department. Professor Sandeep Sukumaran from the Centre for Atmospheric Sciences added: "By combining the accuracy of traditional physics-based models with the speed and efficiency of machine learning, these systems offer exciting new possibilities for early warning and climate adaptation."
&w=3840&q=100)
Business Standard
a day ago
- Science
- Business Standard
AI transforming meteorological prediction through innovation: IIT Delhi
Artificial intelligence is transforming meteorological prediction through innovative approaches to tropical cyclone tracking and monsoon forecasting, two recent studies by the Indian Institute of Technology (IIT), Delhi, researchers have demonstrated. The studies conducted under Professors Sandeep Sukumaran and Hairprasad Kodamana, have been published in the "Journal of Geophysical Research: Machine Learning and Computation". The first study achieved significant advances in monsoon prediction through the application of transformer neural networks. The research team trained their model on a quarter-century of high-resolution satellite precipitation data, enabling the system to accurately forecast monsoon intraseasonal oscillation patterns with an 18-day lead time. "This represents a substantial improvement over existing dynamical models while requiring dramatically fewer computational resources. The AI system's ability to reliably predict active and break phases of the monsoon could have profound implications for agricultural planning and water resource management across South Asia," said PhD scholar KM Anirudh. For the second study, the researchers conducted a comprehensive evaluation of four leading AI weather prediction systems. The research team compared the performance of GraphCast, PanguWeather, Aurora and FourCastNet against conventional numerical weather prediction models. "The AI systems demonstrated remarkable capability in 96-hour cyclone track forecasting, maintaining positional accuracy within 200 kilometres while completing computations in seconds rather than hours. "The Aurora model emerged as the top performer, with researchers attributing its superior performance to the system's transformer-based architecture and incorporation of diverse meteorological datasets," said PhD scholar Pankaj Lal Sahu. "Notably, these AI models successfully internalised complex atmospheric dynamics, including vorticity patterns and pressure gradients, without explicit programming of physical equations, achieving this through advanced machine learning techniques alone," he said. Hariprasad Kodamana, Associate Professor at the Department of Chemical Engineering, informed that the two studies collectively highlight the transformative potential of artificial intelligence in weather prediction. "As extreme weather events become more frequent due to climate change, such AI-powered forecasting tools may prove indispensable for vulnerable communities worldwide," he said. Sandeep Sukumaran, Associate Professor at the Centre for Atmospheric Sciences, explained that by combining the accuracy of traditional physical models with the speed and efficiency of machine learning, these systems offer new possibilities for early warning and climate adaptation. "The research underscores the importance of continued innovation in model architectures and training methodologies to further improve prediction capabilities while maintaining scientific rigour," he added.
