Sea surface temperature (SST) and mixed layer depth (MLD) are crucial in the generation and sustenance of monsoon low-pressure systems in the Bay of Bengal (BoB), driven by complex atmosphere–ocean interactions consequently affecting the Indian Summer Monsoon Rainfall (ISMR). Accurate forecasting of these parameters is vital for understanding and predicting extreme weather events, yet traditional dynamical models often fail to predict them precisely. To address this, we present a deep-learning model based on ConvLSTM to forecast SST and MLD in the BoB for up to four weeks ahead. Wind, freshwater influx, heat fluxes, and ocean stratification are key parameters that control MLD and SST modulations. 20 years of high-resolution data, our model outperforms traditional methods in terms of the biases, correlation coefficient, MAPE, RMSE, and F1 score. SST prediction errors consistently remain quite low, while MLD predictions capture accurate and complex patterns in the Bay, demonstrating the model’s exceptional accuracy. The SHAP analysis provides a clear understanding of the contribution of each input variable to the forecasting of MLD and SST. Overall, this AI model contributes to improved ocean variable forecasting and demonstrates potential for ISMR prediction and disaster management.

Considering turbulence is crucial to understanding clouds. However, covering all scales involved in the turbulent mixing of clouds with their environment is computationally challenging, urging the development of simpler models to represent some of the processes involved. By using full direct numerical simulations as a reference, this study compares several statistical approaches for representing small-scale turbulent mixing. All models use a comparable Lagrangian representation of cloud microphysics, and simulate the same cases of cloud-edge mixing, covering different ambient humidities and turbulence intensities. It is demonstrated that all statistical models represent the evolution of thermodynamics successfully, but not all models capture the changes in cloud microphysics (cloud droplet number concentration, droplet mean radius, and spectral width). Implications of these results for using the presented models as subgrid-scale schemes are discussed.

Aerosol hygroscopicity plays a crucial role in different atmospheric processes, influencing particle size distribution, local visibility, cloud formation, precipitation patterns etc. Hygroscopicity determines the ability of water uptake by the particles. It depends on the particles’ chemical composition. This study presents the importance of hygroscopicity (κ) in aerosols activation process through in-situ ambient measurements and direct numerical simulations (DNS). In-situ measurements were carried out over a one-month period (15 November to 15 December, 2019) by combining two instruments: Humidified Tandem Differential Mobility Analyzer (HTDMA) and Cloud Condensation Nuclei Counter (CCNC) at High Altitude Cloud Physics Laboratory (HACPL). The results from these observations indicate that for particles with diameters 50 nm & 75 nm, activation fraction (AF) increases sharply as κ increases in all supersaturation levels (except at 0.1%). For 110 nm & 150 nm sized particles, a similar trend was found only at lower supersaturation up to 0.3%. Thus, hygroscopicity influences the activation of Aitken mode particles. In the accumulation mode, where particle size primarily governs activation process, hygroscopicity still exerts a significant effect under lower supersaturation conditions. We identified a supersaturation vs size regime wherein hygroscopicity is a decisive factor in aerosol activation. Therefore, the inclusion of κ is essential in numerical models. Implementation of size-dependent κ distribution instead of fixed κ in DNS improves CCN prediction. Modelled activation fraction is within the standard deviation of observation for SS range of 0.3% to 0.7%.

This study presents a paradigm shift by using Deep Neural Networks (DNNs) demonstrating superiority over the traditional methods like Kriging for station-specific precipitation approximation. A thorough analysis of identifying the best nearest neighbour approximation and computation time is carried out to ascertain the computational and methodological supremacy. We propose two innovative NN architectures: one utilizing precipitation, elevation, and location, and the other incorporating additional meteorological parameters like humidity, temperature, and wind speed. Trained on a vast data (1980-2019), these models outperform Kriging across various evaluation metrics (correlation coefficient, root mean square error, bias, and skill score) on a five-year validation set for any given location. This compelling evidence demonstrates the transformative power of deep learning for spatial prediction, offering a robust and precise alternative for hyperlocal precipitation estimation.

Accurate precipitation estimates at individual locations are crucial for weather forecasting and spatial analysis. This study presents a paradigm shift by leveraging Deep Neural Networks (DNNs) to surpass traditional methods like Kriging for station-specific precipitation approximation. We propose two innovative NN architectures: one utilizing precipitation, elevation, and location, and another incorporating additional meteorological parameters like humidity, temperature, and wind speed. Trained on a vast dataset (1980-2019), these models outperform Kriging across various evaluation metrics (correlation coefficient, root mean square error, bias, and skill score) on a five-year validation set. This compelling evidence demonstrates the transformative power of deep learning for spatial prediction, offering a robust and precise alternative for station-specific precipitation estimation.

We ask whether cloud turbulence differs from typical fluid turbulence, and answer in the affirmative: small‐scale turbulence is significantly enhanced in clouds by inertial droplets. This enhancement stems from a mismatch in time scales between vortex evacuation and condensation. Inertial droplets are centrifuged out of vortical regions much faster than condensation can occur. They then grow by condensation, releasing latent heat locally. Our simulations reveal that vortical regions co‐locate with cooler temperatures and higher moisture, leading to sharp density gradients at their boundaries. These gradients drive baroclinic torque, generating small‐scale turbulence. The time scales involved align with our scaling estimates. While our earlier two‐dimensional model anticipated this effect, it underestimated its intensity and limited it to sub‐Kolmogorov scales. The enhancement of turbulence over inertial scales that we find here could drive rapid raindrop growth.

Droplet growth and size spectra play a crucial role in the microphysics of atmospheric clouds. However, it is challenging to represent droplet growth rate accurately in cloud-resolving models such as Large Eddy Simulations (LESs). The assumption of "well-mixed" condition within each grid cell, often made by traditional LES solvers, typically falls short near the edges of clouds, where sharp gradients in water vapor supersaturation occur. This under-resolution of supersaturation gradients can lead to significant errors in prediction of droplet growth rate, which in turn affects the prediction of buoyancy at cloud edges, as well as forecast of precipitation. In "superdroplet" based LES model, a Lagrangian coarse-graining approach groups multiple droplets into superdroplets, each encompassing a specific number and size of actual droplets. The superdroplets are advected by the underlying LES velocity field, and the growth rate of these superdroplets is based on the filtered supersaturation field represented by the LES. To overcome the limitations of the "well-mixed" assumption, we propose a parameterization for superdroplet growth using high-fidelity Direct Numerical Simulation (DNS) data. We introduce a novel clustering algorithm to map droplets in DNS fields to superdroplets. The effective supersaturation at each superdroplet location is computed by averaging the unfiltered supersaturation of the associated droplets, which may differ from the value of filtered supersaturation at the superdroplet location. We then develop a machine learning-based parameterization to relate the effective growth rate of superdroplets to other filtered DNS flow variables. Preliminary results show a promising R^2 value of nearly 0.9 between the predicted and true effective supersaturation values for the superdroplets, for a range of superdroplet multiplicities.

Simulated data of atmospheric variables like precipitation is very important for climate science research, especially for understanding future scenarios. Such data is generated by global or regional climate models (GCM/RCM), for both past and future periods under different initial conditions. Unfortunately, these models frequently display biases in simulated precipitation data compared to ground observations, due to their inability to incorporate the entire physics of the process accurately. It is imperative to mitigate such biases using contemporary techniques to make the simulated data a valuable end product. Our aim in this research is to correct seasonal (from June to September) Indian Summer Monsoon Rainfall (ISMR) data as simulated by the Climate Forecast System (CFS) over the Indian subcontinent, with the Global Precipitation Climatology Project (GPCP) data serving as the ground truth reference. This period is important as more than one-third of India's annual rainfall occurs during these months. This study presents a novel Deep Learning (DL) based architecture known as the Convolutional Neural Network for Bias Correction (CNNBC) that aims to calibrate the model-simulated data with past observations, to address these biases while preserving the statistical properties and spatial correlations relationships among grid points and enhancing the spatial mean of precipitation estimates. To evaluate the effectiveness of our proposed method, we compare its performance to that of three other statistical and DL techniques: quantile mapping (QM), quantile delta mapping (QDM), and the SRDRN model. The comparative analysis demonstrates that the CNNBC model outperforms the other in terms of our task.

Cloud droplet dynamics is an important part of cloud physics. This element of cloud physics analyses the features of each droplet, including its size distribution, probability density and mean saturation. The cloud's structure is significantly important for the Earth's atmosphere and this structure is affected by changes in the droplet's micro-physical properties. In order to investigate and understand the dynamics of cloud droplets in both the high and low vortex areas, data obtained from Direct Numeric Simulations (DNS) are utilized. Data generated from simulations of cumulus clouds, which are defined as low-level clouds located between 800 and 1200 m above the surface of the earth. DNS data reveals complex droplet dynamics on a scale that is three-dimensional. When employing conventional machine learning methods, the processing of data relating to dynamic droplets requires a substantial amount of CPU resources. In this study, we discussed the advantages of using quantum mechanisms in cloud physics in order to investigate the complicated nature of cloud droplets. The use of quantum computing in the study of droplet dynamics using the quantum k-mean approach was further investigated in the discussion. Quantum machine learning is used to study the micro-physical characteristics of cloud droplets in order to investigate the effect that droplet dynamics have on the overall structure of clouds. The current topic of discussion delves more into the specifics of how data relating to DNS can be processed by an analog quantum computer in order to deal with enormous amounts of data in this specific area of research.

Accurate near real time precipitation forecasting has several benefits, including water resource management, dam discharge and flash flood management. In this regard, deep-learning offers good value in precipitation now-casting, particularly when supplemented with reliable observational records, e.g. Radar images. This study employs deep learning (DL) models for precipitation now-casting utilizing Radar precipitation data over Bhopal city located in central India and tests its efficacy during the monsoon (JJAS) 2021 season, with a 20-min temporal resolution. Out of the three methods tested for forecasting, the DL model ConvLSTM outperforms ConvGRU model, and persistence baseline method, in terms of spatial and temporal correlation, skill score, and RMSE, and is thus chosen for further investigations. The ConvLSTM model provides an accuracy of up to 75% for the 1st lead time step forecast and gradually decreases for further time steps going down to approximately 35% at the 5th lead time step forecast. Moreover, while comparing directly from ground truth, the model is able to capture the temporal (sequential) linkage in data. The findings show that deep-learning-based models have the potential to improve precipitation now-casting.

Direct Numerical Simulation (DNS) of turbulent cloud parcels are presented where particles evolve in response to the local values in supersaturation (s) led by turbulent fluctuations. A pseudo-spectral DNS is modified to incorporate aerosol particles and Cloud Condensation Nuclei (CCN) activation applying a droplet growth equation based on κ-Köhler theory that works for the whole range of warm cloud particles, from un-activated deliquesced aerosol particles to activated cloud droplets, that grow by condensation. The Lagrangian microphysics applied ensures that each particle experiences a supersaturation (fluctuations added to the mean) corresponding to its near-particle gas phase surroundings, which differs from the uniform parcel view where an average value of supersaturation, s¯, is applied to every particle in the domain. A non-turbulent Lagrangian parcel model with similar mean thermodynamics and CCN properties is compared with turbulent DNS parcels of varying fluctuation intensities to distinguish the impact of turbulence on CCN activation. It is shown that, in a given distribution, a subset of aerosols respond to fluctuations rather than the mean thermodynamics allowing particles of similar dry size to co-exist as un-activated haze particles as well as cloud droplets. The cloud microphysical properties are analysed in a series of DNS experiments with contrasting thermodynamic and aerosol properties. DNS results agrees with the general understanding of aerosol activation and cloud droplets evolution in response to various background conditions such as pristine and polluted aerosol distributions, composition of CCN with respect to organic - inorganic components and magnitude of vertical velocity. DNS considers turbulence-microphysics interactions in such problems providing additional knowledge on the role of turbulence in clouds. It is argued that incorporating turbulent fluctuations in simulating the CCN activation and droplet growth is well reasoned and a constructive way forward.

Deep Learning (DL) based downscaling has become a popular tool in earth sciences recently. Increasingly, different DL approaches are being adopted to downscale coarser precipitation data and generate more accurate and reliable estimates at local ( few km or even smaller) scales. Despite several studies adopting dynamical or statistical downscaling of precipitation, the accuracy is limited by the availability of ground truth. A key challenge to gauge the accuracy of such methods is to compare the downscaled data to point-scale observations which are often unavailable at such small scales. In this work, we carry out the DL-based downscaling to estimate the local precipitation data from the India Meteorological Department (IMD), which was created by approximating the value from station location to a grid point. To test the efficacy of different DL approaches, we apply four different methods of downscaling and evaluate their performance. The considered approaches are (i) Deep Statistical Downscaling (DeepSD), augmented Convolutional Long Short Term Memory (ConvLSTM), fully convolutional network (U-NET), and Super-Resolution Generative Adversarial Network (SR-GAN). A custom VGG network, used in the SR-GAN, is developed in this work using precipitation data. The results indicate that SR-GAN is the best method for precipitation data downscaling. The downscaled data is validated with precipitation values at IMD station. This DL method offers a promising alternative to statistical downscaling.

The prediction of surface ozone is essential attributing to its impact on human and environmental health. Volatile organic compounds (VOCs) are crucial in driving ozone concentration; particularly in urban areas where VOC limited regimes are prominent. The limited measurements of VOCs, however, hinder assessing the VOC-ozone relationship. This work applies machine learning (ML) algorithms for temporal forecasting of surface ozone over a metropolitan city in India. The availability of continuous VOCs measurement data along with meteorology and other pollutants during 2014-2016 makes it possible to deduce the influence of various input parameters on surface ozone prediction. After evaluating the best ML model for ozone prediction, simulations were carried out using varied input combinations. The combination with isoprene, meteorology, NOx, and CO (Isop + MNC) was the best with RMSE 4.41 ppbv and MAPE 6.77%. A season-wise comparison of simulations having all data, only meteorological data and Isop + MNC as input showed that Isop + MNC simulation gives the best results during the summer season (RMSE: 5.86 ppbv, MAPE: 7.05%). This shows the increased ability of the model to capture ozone peaks (high ozone during summer) relatively better when isoprene data is used. The overall results highlight that using all available data doesn't necessarily give best prediction results; also critical thinking is essential when evaluating the model results.

Deep learning (DL), a potent technology to develop Digital Twin (DT), for weather prediction using cubed spheres (DLWP-CS) was recently proposed to facilitate data-driven simulations of global weather fields. DLWP-CS is a temporal mapping algorithm wherein time-stepping is performed through U-NET. Although DLWP-CS has shown impressive results for fields, such as temperature and geopotential height, this technique is complicated and computationally challenging for a complex, non-linear field, such as precipitation, which depends on other prognostic environmental co-variables. To address this challenge, we modify the DLWP-CS and call our technique “modified DLWP-CS” (MDLWP-CS). In this study, we transform the architecture from a temporal to a spatio-temporal mapping (multivariate setup), wherein precursor(s) of precipitation can be used as input. As a proof of concept, as a first simple case, a 2-m surface air temperature is used to predict precipitation using MDLWP-CS. The model is trained using hourly ERA-5 reanalysis and the resulting experimental findings are compared to two benchmark models, viz, the linear regression and an operational numerical weather prediction model, which is the Global Forecast System (GFS). The fidelity of MDLWP-CS is much better compared to linear regression and the results are equivalent to GFS output in terms of daily precipitation prediction with 1 day lag. These results provide an encouraging framework for an efficient DT that can facilitate speedy, high fidelity precipitation predictions.

Climate change and human activity have increased fires in India. Fine particulate matter ( PM_2.5 ) is released into the atmosphere by stubble burning in Punjab and Haryana and forest fires in the north-eastern and central areas of the country. Accurate short-term PM_2.5 estimates are essential to protect human health and reduce acute air pollution. However, global air quality forecasting methods grapple with a persistent assumption of fire emissions. They use near-real-time fire emissions throughout the prediction cycle. Air quality forecasts are prone to inaccuracies and biases due to fire emissions’ dynamic nature. We employ spatiotemporal deep learning techniques, specifically ConvLSTM and ConvGRU, to forecast fire emission locations up to three days in advance. Through our evaluation, we find that ConvLSTM outperforms ConvGRU in terms of prediction accuracy and performance. The chosen model provides a very good correlation coefficient ( ≈ 0.8 ) for the 1st day forecast and a moderate value (0.5 - 0.55) for subsequent 2nd and 3rd days forecasts. The predictors NDVI, temperature, wind, surface pressure, and total cloud cover are included to our model training to improve these correlations. In Punjab-Haryana, wind input improves results. This fire burning location prediction method could improve air quality forecasting. Our deep learning model can improve forecasts by revealing the complex interactions of components and reflecting fire emissions’ dynamic nature. This research may help improve air quality forecasts in the face of rising fire events, protecting communities across the Indian subcontinent.

The analysis and investigation of the data obtained from Direct Numerical (DNS) simulation of droplet dynamics in cloud turbulence is a complex and time-consuming task when performaed on traditional computers. The DNS data generally have, a high spatial resolution $\approx 1mm$ and require considerable space to store. It is tedious to find specific features of this data, such as locating high and low vortex areas in cloud turbulence using machine learning algorithms. In this research, we employ quantum computing to examine and analyze cloud droplet dynamics data and present a quantum supervised machine learning algorithm, namely, a support vector machine (SVM) to segregate low and high vortex regions and investigate the droplet characteristics in those regions. The result show that use of quantum computers can accelerate the entire process, and quantum mechanics tools, such as quantum kernels and quantum circuits can better manage the complex nature of data than traditional methods.

Lightning strikes are a well-known danger, and are a leading cause of accidental fatality worldwide. Unfortunately, lightning hazards seldom make headlines in international media coverage because of their infrequency and the low number of casualties each incidence. According to readings from the TRMM LIS lightning sensor, thunderstorms are more common in the tropics while being extremely rare in the polar regions. To improve the precision of lightning forecasts, we develop a technique similar to LightNet's, with one key modification. We didn't just base our model off the results of preliminary numerical simulations; we also factored in the observed fields' time-dependent development. The effectiveness of the lightning forecast rose dramatically once this adjustment was made. The model was tested in a case study during a thunderstorm. Using lightning parameterization in the WRF model simulation, we compared the simulated fields. As the first of its type, this research has the potential to set the bar for how regional lightning predictions are conducted in the future because of its data-driven approach. In addition, we have built a cloud-based lightning forecast system based on Google Earth Engine. With this setup, lightning forecasts over West India may be made in real time, giving critically important information for the area.

This paper examines the impact of cloud-base turbulence on activation of cloud condensation nuclei (CCN). Following our previous studies, we contrast activation within a nonturbulent adiabatic parcel and an adiabatic parcel filled with turbulence. The latter is simulated by applying a forced implicit large-eddy simulation within a triply periodic computational domain of 64(3) m(3). We consider two monodisperse CCN. Small CCN have a dry radius of 0.01 mu m and a corresponding activation (critical) radius and critical supersaturation of 0.6 mu m and 1.3%, respectively. Large CCN have a dry radius of 0.2 mu m and feature activation radius of 5.4 mu m and critical supersaturation 0.15%. CCN are assumed in 200-cm(-3) concentration in all cases. Mean cloud-base updraft velocities of 0.33, 1, and 3 m s(-1) are considered. In the nonturbulent parcel, all CCN are activated and lead to a monodisperse droplet size distribution above the cloud base, with practically the same droplet size in all simulations. In contrast, turbulence can lead to activation of only a fraction of all CCN with a nonzero spectral width above the cloud base, of the order of 1 mu m, especially in the case of small CCN and weak mean cloud-base ascent. We compare our results to studies of the turbulent single-size CCN activation in the Pi chamber. Sensitivity simulations that apply a smaller turbulence intensity, smaller computational domain, and modified initial conditions document the impact of specific modeling assumptions. The simulations call for a more realistic high-resolution modeling of turbulent cloud-base activation.