KHKieran Mark Rainwater Hunt

Kieran Mark Rainwater Hunt

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Kieran Mark Rainwater Hunt has a total of 162 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Evaluating Machine‐Learning Weather Prediction for UK Windstorm Weather Warnings

    by James Hewitt, Ambrogio Volonté, Ben Harvey, Andressa Andrade Cardoso, Kieran M. R. Hunt, Natalie J. Harvey, Oscar Martinez‐Alvarado, Suzanne L. Gray, Helen F. Dacre, Todd R. Jones

    Machine‐learning weather prediction (MLWP) has emerged as a powerful new forecasting tool, with the potential to complement or even replace numerical weather prediction (NWP) in early warning systems. MLWP models are computationally cheap once trained, offering more timely, accessible, and energy‐efficient forecasts than NWP. Yet their ability to capture weather extremes, particularly strong winds associated with extra‐tropical cyclones, remains insufficiently understood. This study provides the first comprehensive evaluation of MLWP performance for dynamically diverse UK windstorms, addressing a critical evidence gap before these models can be responsibly deployed in weather risk management. We evaluate eight state‐of‐the‐art MLWP models against reanalysis, NWP analysis, and an operational NWP ensemble for six contrasting UK cyclones from winter 2023/24. We find substantial inter‐storm and inter‐model variability in forecast skill at 24‐h lead time. Forecast accuracy is found to be linked to the underlying atmospheric dynamics and spatial scales of the wind‐producing features: wind maxima associated with mesoscale structures are systematically underestimated across all MLWP models. Significant variability is also found in the representation of mesoscale storm structure, and the highest‐resolution MLWP model (Aurora‐0.1) does not deliver uniform improvements across the cyclone set. In‐depth analysis of two mature Shapiro–Keyser cyclones reveals a consistent tendency for MLWP models to smooth sharp frontal gradients and misrepresent key mesoscale features, including bent‐back warm fronts and frontal‐fracture regions. These deficiencies are physically consistent with the underestimation of localised extreme winds. Overall, our findings highlight the current limitations of MLWP for underpinning UK wind warnings. While these models are not yet suitable to replace NWP for hazardous windstorm prediction entirely, their demonstrated skill in capturing storm tracks and synoptic‐scale evolution points to promising complementary roles in future national weather forecasting and resilience frameworks.

  • Open Access

    Stop Using Root-Mean-Square Error as a Precipitation Target!

    by Kieran M. R. Hunt

    Root-mean-square error (RMSE) remains the default training loss for data-driven precipitation models, despite precipitation being semicontinuous, zero inflated, strictly nonnegative, and heavy tailed. This Gaussian-implied objective misspecifies the data-generating process because it tolerates negative predictions, underpenalizes rare heavy events, and ignores the mass at zero. We suggest replacing RMSE with the Tweedie deviance, a likelihood-based and differentiable loss from the exponential–dispersion family, with variance function V ( μ ) = μ p . For 1 < p < 2, it yields a compound Poisson–gamma distribution with a point mass at zero and a continuous density over y > 0, matching observed precipitation characteristics. In this paper, we (i) estimate p from the variance–mean power law and show that precipitation across temporal aggregations is far from Gaussian and that the Tweedie power p increases with accumulation length toward a gamma limit and (ii) demonstrate consistent skill gains when training deep data-driven models with Tweedie deviance instead of RMSE. In diffusion-model downscaling over Beijing, Tweedie loss improves wet-pixel MAE and extreme recall (∼0.611 vs 0.566 at the 99th percentile). In convolutional long short-term memory (ConvLSTM) nowcasting over Kolkata, Tweedie loss yields improvements in wet-pixel MAE and dry-pixel hit rates, with improvements autoregressively compounding over lead time (for MAE, from 3.5% at t + 1 growing to 8.7% at t + 4). Extreme recall at the 99.9th percentile is not significantly higher at t + 1 but becomes significantly higher by t + 3 and t + 4. Because the Tweedie deviance is continuous in p , it adapts smoothly across scales, offering a statistically justified, practical replacement for RMSE in precipitation-based learning tasks.


  • Open Access

    Reply on CC1

    by Karan Ruparell, Dai Yamazaki, Kieran Hunt, Hannah Cloke, Christel Prudhomme, Florian Pappenberger, Matthew Chantry

    Reservoirs fundamentally alter downstream river flow regimes, decoupling discharge from natural meteorological forcing and challenging standard hydrological prediction. While data-driven models, such as Long Short-Term Memory (LSTM) networks, show promise in regulated catchments, it remains unclear how training data composition across natural and regulated rivers influences model generalisability and behaviour. In this study, we investigate how the presence or absence of reservoir-influenced catchments in training data impacts model performance across different flow regimes and alters the physical drivers the models learn to rely on. Using carefully matched subsets of the CAMELS-GB dataset, we trained separate specialist LSTMs (reservoir and non-reservoir), a pooled Full LSTM, and a multi-headed Hydra-LSTM to investigate whether explicit architectural specialisation offers any advantage over pooled training alone. Models were evaluated on held-out test gauges using standard performance metrics and gradient importance analysis to interpret feature reliance. Our results demonstrate that exposure to reservoir-influenced catchments during training is essential. Models trained exclusively on natural catchments consistently overestimate the mean and variance of regulated flows. Conversely, training exclusively on reservoir-influenced data degrades performance on non reservoir-influenced rivers (KGE reduction of ≥ 0.1) giving importance primarily to anthropogenic static features, such as abstraction rates, at the expense of precipitation drivers. A single Full LSTM trained on combined data matched the performance of both specialist models in their respective domains, implicitly switching its feature reliance between regimes. The Hydra-LSTM performed comparably to the Full LSTM throughout, indicating that the shared body may act as a regulariser limiting over-specialisation, but that explicit architectural specialisation provides no further benefit under these conditions. We conclude that pooling training data across regimes is a highly effective strategy for general-purpose modelling. However, case studies highlight a fundamental limitation: purely meteorological inputs remain insufficient for predicting flows in heavily managed single-purpose reservoirs, where unobserved human operational decisions dominate the hydrograph.

  • Open Access

    Reply on RC2

    by Karan Ruparell, Dai Yamazaki, Kieran Hunt, Hannah Cloke, Christel Prudhomme, Florian Pappenberger, Matthew Chantry

    Reservoirs fundamentally alter downstream river flow regimes, decoupling discharge from natural meteorological forcing and challenging standard hydrological prediction. While data-driven models, such as Long Short-Term Memory (LSTM) networks, show promise in regulated catchments, it remains unclear how training data composition across natural and regulated rivers influences model generalisability and behaviour. In this study, we investigate how the presence or absence of reservoir-influenced catchments in training data impacts model performance across different flow regimes and alters the physical drivers the models learn to rely on. Using carefully matched subsets of the CAMELS-GB dataset, we trained separate specialist LSTMs (reservoir and non-reservoir), a pooled Full LSTM, and a multi-headed Hydra-LSTM to investigate whether explicit architectural specialisation offers any advantage over pooled training alone. Models were evaluated on held-out test gauges using standard performance metrics and gradient importance analysis to interpret feature reliance. Our results demonstrate that exposure to reservoir-influenced catchments during training is essential. Models trained exclusively on natural catchments consistently overestimate the mean and variance of regulated flows. Conversely, training exclusively on reservoir-influenced data degrades performance on non reservoir-influenced rivers (KGE reduction of ≥ 0.1) giving importance primarily to anthropogenic static features, such as abstraction rates, at the expense of precipitation drivers. A single Full LSTM trained on combined data matched the performance of both specialist models in their respective domains, implicitly switching its feature reliance between regimes. The Hydra-LSTM performed comparably to the Full LSTM throughout, indicating that the shared body may act as a regulariser limiting over-specialisation, but that explicit architectural specialisation provides no further benefit under these conditions. We conclude that pooling training data across regimes is a highly effective strategy for general-purpose modelling. However, case studies highlight a fundamental limitation: purely meteorological inputs remain insufficient for predicting flows in heavily managed single-purpose reservoirs, where unobserved human operational decisions dominate the hydrograph.

  • Open Access

    Reply on RC1

    by Karan Ruparell, Dai Yamazaki, Kieran Hunt, Hannah Cloke, Christel Prudhomme, Florian Pappenberger, Matthew Chantry

    Reservoirs fundamentally alter downstream river flow regimes, decoupling discharge from natural meteorological forcing and challenging standard hydrological prediction. While data-driven models, such as Long Short-Term Memory (LSTM) networks, show promise in regulated catchments, it remains unclear how training data composition across natural and regulated rivers influences model generalisability and behaviour. In this study, we investigate how the presence or absence of reservoir-influenced catchments in training data impacts model performance across different flow regimes and alters the physical drivers the models learn to rely on. Using carefully matched subsets of the CAMELS-GB dataset, we trained separate specialist LSTMs (reservoir and non-reservoir), a pooled Full LSTM, and a multi-headed Hydra-LSTM to investigate whether explicit architectural specialisation offers any advantage over pooled training alone. Models were evaluated on held-out test gauges using standard performance metrics and gradient importance analysis to interpret feature reliance. Our results demonstrate that exposure to reservoir-influenced catchments during training is essential. Models trained exclusively on natural catchments consistently overestimate the mean and variance of regulated flows. Conversely, training exclusively on reservoir-influenced data degrades performance on non reservoir-influenced rivers (KGE reduction of ≥ 0.1) giving importance primarily to anthropogenic static features, such as abstraction rates, at the expense of precipitation drivers. A single Full LSTM trained on combined data matched the performance of both specialist models in their respective domains, implicitly switching its feature reliance between regimes. The Hydra-LSTM performed comparably to the Full LSTM throughout, indicating that the shared body may act as a regulariser limiting over-specialisation, but that explicit architectural specialisation provides no further benefit under these conditions. We conclude that pooling training data across regimes is a highly effective strategy for general-purpose modelling. However, case studies highlight a fundamental limitation: purely meteorological inputs remain insufficient for predicting flows in heavily managed single-purpose reservoirs, where unobserved human operational decisions dominate the hydrograph.

  • Open Access

    Synoptic and large‐scale drivers of extreme precipitation events in the Himalayas

    by Priya Bharati, Pranab Deb, Kieran M. R. Hunt

    Extreme precipitation events (EPEs) in the Western Himalayas (WH) have significant cross‐sector impacts, yet understanding of their seasonal characteristics and synoptic drivers remains inadequate. This study presents an analysis of EPEs across the full annual cycle of the Himalayas from 1979 to 2023, identifying their spatial and seasonal distributions and associated precursor weather regimes. Based on the frequency of seasonal EPEs, the Himalayas are categorized into winter‐dominated (Karakoram, WH, and northwest Central Himalayas [CH]) and summer‐dominated (CH to Eastern Himalayas [EH]). By clustering circulation fields during EPEs, six dominant weather regimes are identified: four monsoon‐driven (Bay of Bengal Low‐Pressure System [BoB‐LPS], Arabian Sea LPS [AS‐LPS], Monsoon‐Break, and Western‐Disturbance Monsoon [WD‐Monsoon]) and two winter‐driven (WH‐WD and CH‐WD). Each regime exhibits a distinct atmospheric circulation pattern. The BoB‐LPS and AS‐LPS regimes are characterized by quasi‐stationary blocking by an anomalous upper‐level anticyclone situated over the WH along with the LPS formed over the BoB and AS, respectively. The Monsoon‐Break cluster reflects a weakened monsoon trough and suppressed convection over the BoB and core monsoon regions. In contrast, the WH‐WD and CH‐WD clusters are associated with upper‐level cyclonic anomalies characteristic of WDs, embedded within broader wave‐train‐like circulation anomalies that are consistent with previously reported North Atlantic Oscillation (NAO)–subtropical jet (STJ)–WD pathways. There is a spatial dipole in composite precipitation: EPEs in monsoon‐driven clusters are associated with anomalous moisture transport from the BoB and AS, which results in dry conditions in the winter‐dominated Himalayas, while EPEs in WD‐driven regimes result from anomalous westerly moisture flow from the AS and associated dry conditions in the summer‐dominated Himalayas. This study offers a synoptic‐scale framework for interpreting the circulation regimes associated with Himalayan EPEs and for identifying conditions that may be favourable for their occurrence.

  • Seasonal predictability of winter PM 2.5 pollution severity in India through a strong pollution–extratropical storm connection

    by Yuanyu Xie, Kieran M. R. Hunt, Rohit Gupta, Gargee Goswami, Mi Zhou, Wenhao Dong, Denise L. Mauzerall

    India suffers from severe fine particulate matter (PM 2.5 ) air pollution in winter that harms public health and disrupts economic activities. Seasonal prediction of pollution severity could facilitate proactive mitigation planning but is hindered by incomplete understanding of the drivers of pollution variations. Here, using over a decade of quality-controlled Indian government PM 2.5 monitoring data, we find that ∼70% of the interannual variations in winter PM 2.5 pollution across northern India are explained by western disturbances (WDs)—prevailing extratropical storms originating from the Mediterranean region—through PM 2.5 removal via enhanced ventilation and precipitation. These WDs are regulated by the location and strength of the subtropical jet traceable to sea surface temperature (SST) anomalies across the North Atlantic and tropical Indian Ocean during the preceding autumn. Using these preseason SST anomalies, we demonstrate the potential to predict winter WD precipitation and PM 2.5 pollution severity one season ahead, with good agreement with observations ( r 2 = 0.52 to 0.69), which could facilitate proactive public policies and stakeholder mitigation actions.

  • Open Access

    Comment on egusphere-2026-2909

    by Karan Ruparell, Dai Yamazaki, Kieran Hunt, Hannah Cloke, Christel Prudhomme, Florian Pappenberger, Matthew Chantry

    Reservoirs fundamentally alter downstream river flow regimes, decoupling discharge from natural meteorological forcing and challenging standard hydrological prediction. While data-driven models, such as Long Short-Term Memory (LSTM) networks, show promise in regulated catchments, it remains unclear how training data composition across natural and regulated rivers influences model generalisability and behaviour. In this study, we investigate how the presence or absence of reservoir-influenced catchments in training data impacts model performance across different flow regimes and alters the physical drivers the models learn to rely on. Using carefully matched subsets of the CAMELS-GB dataset, we trained separate specialist LSTMs (reservoir and non-reservoir), a pooled Full LSTM, and a multi-headed Hydra-LSTM to investigate whether explicit architectural specialisation offers any advantage over pooled training alone. Models were evaluated on held-out test gauges using standard performance metrics and gradient importance analysis to interpret feature reliance. Our results demonstrate that exposure to reservoir-influenced catchments during training is essential. Models trained exclusively on natural catchments consistently overestimate the mean and variance of regulated flows. Conversely, training exclusively on reservoir-influenced data degrades performance on non reservoir-influenced rivers (KGE reduction of ≥ 0.1) giving importance primarily to anthropogenic static features, such as abstraction rates, at the expense of precipitation drivers. A single Full LSTM trained on combined data matched the performance of both specialist models in their respective domains, implicitly switching its feature reliance between regimes. The Hydra-LSTM performed comparably to the Full LSTM throughout, indicating that the shared body may act as a regulariser limiting over-specialisation, but that explicit architectural specialisation provides no further benefit under these conditions. We conclude that pooling training data across regimes is a highly effective strategy for general-purpose modelling. However, case studies highlight a fundamental limitation: purely meteorological inputs remain insufficient for predicting flows in heavily managed single-purpose reservoirs, where unobserved human operational decisions dominate the hydrograph.

  • Open Access

    Kilometer‐Scale (1‐km) Convection‐Permitting Simulations Over the Hindu Kush‐Himalayan Region: Processes of Winter Precipitation

    by S. Nischal, Raju Attada, Kieran M. R. Hunt, Chandrasekar Radhakrishnan, Valentine Anantharaj

    Kilometer‐scale (k‐scale) simulations, with explicit treatment of convection at sub‐grid scales, are useful for understanding precipitation characteristics. Such simulations with their high spatiotemporal resolution can be particularly valuable in complex topographic regions, such as the Hindu Kush Himalayas (HKH), where sparse observations and uncertainties in coarse‐resolution data sets pose challenges. This study evaluates an AMIP style k‐scale (1 km) simulation, initialized from the ECMWF IFS analysis, for winter precipitation (December 2018–February 2019) over the HKH region, with a focus on mean and extreme characteristics, elevation‐dependent variability, and diurnal precipitation cycles. The kilometer‐scale simulations were validated using available station observations and multiple high‐resolution gridded precipitation data sets, including gauge‐based, satellite‐derived, and reanalysis products. The model shows high fidelity to the observed station‐based precipitation magnitudes and realistically depicts the spatial distribution of precipitation, particularly the ridge‐valley variations, often missed in coarser products. In general, it aligns more with reanalysis data sets and closely matches station observations as well. Mean precipitation exhibits sensitivity to elevation, and the highest rates occur at about 2,500 m in most of the reference products (observations/reanalysis), which the k‐scale model represents well. The diurnal cycle depicts sub‐daily precipitation maxima in the local afternoon and early morning hours. The analysis for precipitation extremes indicates the model's close fidelity with reanalysis products in capturing higher‐intensity and prolonged precipitation events in the western Himalayas. Radiosonde profiles and atmospheric thermodynamic characteristics highlight a highly saturated and unstable environment during extremes, which is favorable for enhanced convective developments and heavy precipitation. The model captures these atmospheric conditions well and represents the localized variations and intensifications in valley wind flows during extremes, which are often missed in coarser‐resolution and parameterized ERA5 data. Our findings highlight the fidelity of this k‐scale AMIP‐style model over coarser‐resolution, parameterized models in resolving subgrid‐scale processes in the complex terrain regions of the HKH.

  • Open Access

    Comment on egusphere-2026-2909

    by Karan Ruparell, Dai Yamazaki, Kieran Hunt, Hannah Cloke, Christel Prudhomme, Florian Pappenberger, Matthew Chantry

    Reservoirs fundamentally alter downstream river flow regimes, decoupling discharge from natural meteorological forcing and challenging standard hydrological prediction. While data-driven models, such as Long Short-Term Memory (LSTM) networks, show promise in regulated catchments, it remains unclear how training data composition across natural and regulated rivers influences model generalisability and behaviour. In this study, we investigate how the presence or absence of reservoir-influenced catchments in training data impacts model performance across different flow regimes and alters the physical drivers the models learn to rely on. Using carefully matched subsets of the CAMELS-GB dataset, we trained separate specialist LSTMs (reservoir and non-reservoir), a pooled Full LSTM, and a multi-headed Hydra-LSTM to investigate whether explicit architectural specialisation offers any advantage over pooled training alone. Models were evaluated on held-out test gauges using standard performance metrics and gradient importance analysis to interpret feature reliance. Our results demonstrate that exposure to reservoir-influenced catchments during training is essential. Models trained exclusively on natural catchments consistently overestimate the mean and variance of regulated flows. Conversely, training exclusively on reservoir-influenced data degrades performance on non reservoir-influenced rivers (KGE reduction of ≥ 0.1) giving importance primarily to anthropogenic static features, such as abstraction rates, at the expense of precipitation drivers. A single Full LSTM trained on combined data matched the performance of both specialist models in their respective domains, implicitly switching its feature reliance between regimes. The Hydra-LSTM performed comparably to the Full LSTM throughout, indicating that the shared body may act as a regulariser limiting over-specialisation, but that explicit architectural specialisation provides no further benefit under these conditions. We conclude that pooling training data across regimes is a highly effective strategy for general-purpose modelling. However, case studies highlight a fundamental limitation: purely meteorological inputs remain insufficient for predicting flows in heavily managed single-purpose reservoirs, where unobserved human operational decisions dominate the hydrograph.