International Project to Study the Environmental Influences on the Lightning of Warm-based Deep Convection

Project leader: Dr V. Phillips, INES, University of Lund, Lund, Sweden

Unfunded collaborator: Professor R. Albrecht, Department of Atmospheric Sciences, University of Sao Paulo (USP), Brazil

Unfunded collaborator: Professor M. Ohlsson, Centre for Environmental and Climate Science, Lund University, Lund, Sweden

Overview of project

A project entitled “Multiple Environmental Influences on the Lightning of Warm-based Deep Convection Simulated with AI” led by Lund University was funded in 2021 by the Swedish Research Council (VEtenskapsrådet ‘VR’; award number 2021-05547). The project lasts 4 years from about 2022 to 2025.

Institutions in USA and Sweden are involved in the collaboration:

· Department of Atmospheric Sciences - Institute Astronômical and Geophysical, University of Sao Paulo in Brazil;

· Centre for Environmental and Climate Science, Lund University, in Sweden;

· Department of Physical Geography and Ecosystem Science, Lund University, in Sweden.

The over-arching aim of the project is to accelerate the computations of the lightning, so as to apply the cloud/electrification model to warm-based thunderstorms. Various scientific hypotheses about how the environment controls the frequency and properties of the lightning will then be tackled.

Background and previous work

Lightning is created by separation of charge in ice-ice collisions in the presence of supercooled cloud-liquid. The extent and polarity of the charging is dependent on the sizes of colliding particles, and on the distribution of mass content of cloud-liquid over subzero temperatures in any storm. Thus, the frequency and properties of lightning (e.g. cloud to ground flashes conveying either negative or positive charge) are related to the microphysical regime of the thunderstorm and to environmental conditions of instability and aerosols.

If the bases of convective clouds are sufficiently warm, then their precipitation is generated by the warm rain process of coalescence. This is more likely when the aerosol conditions are more nearly maritime, with fewer aerosol particles for initiation of cloud-droplets. Then ice precipitation tends to form by raindrop-freezing, resulting in high-density graupel or hail. In such warm-based clouds, the removal of cloud-liquid by warm rain during lengthy ascent from the base to the freezing level then causes weaker amounts of supercooled liquid aloft. The lack of supercooled cloud-liquid then reverses the polarity of much of the charging relative to colder-based storms, so their charge structure is more likely to be ‘normal’ (see schematic diagram below). Consequently, there are reasons to expect warm-based clouds to have different lightning properties compared to cold-based clouds simulated in our previous study, as elucidated in previous work (Phillips et al. 2020, 2022).

Conceptual picture of how the environment factors of aerosol conditions (cloud condensation nuclei: “CCN”; ice nuclei: “IN”) and thermodynamics (“ascent”, “cloud-base”) influence the charge structure in a thunderstorm. Charging in graupel-crystal collisions occurs in a phase-space of liquid water content and temperature (central panel), with regimes of negative and positive charging. The location of most of the charging in this phase-space (long slanted ellipses in central panel) determines the overall charge structure of normal and inverted/anomalous storms. Most storms are ‘normal’ in charge structure (left panels) as they involve predominant positive charging of graupel (shaded ellipses), which tends to fall out, and the resulting net negative charge on the snow/crystals (shaded stars) of the storm causes negative charge to go to ground in cloud-to-ground flashes. Vice versa for ‘inverted’ charge structure and positive charge to ground (right panels). The fraction of all charge separated in the storm that is positive for graupel/hail is . From Phillips et al. (2022).

Lightning modeling and AI

A lightning scheme was implemented into our aerosol-cloud model (AC) by Phillips et al. (2020, 2022). AC is briefly described on other webpages here. This lightning scheme involved a modification of a scheme by Barthe et al. (2005). Charge is separated in ice-ice collisions involving all the microphysical species of ice using a novel interpretation of results from Takahashi in the 1970s. We included a dependency of charge separated on ice morphology and size. The electric field is solved, yielding trigger-points. From each trigger-point, a bidirectional leader channel is traced exactly parallel and anti-parallel to the field vector (see diagram below).

An iterative scheme combined with a fractal algorithm for the number density of branches determines the 3D network of branches connected to the leader in the charged part of the cloud. Neutralisation of local charge is then evaluated. It is this ‘branching scheme’ that is computationally expensive.

$Text Box: Schematic diagram of the fractal branching algorithm as applied to a single flash in the scheme by Barthe et al. (2005), (from Phillips et al. 2020). Two leaders propagate from the trigger-point. Each hemispheric shell has a number of branches according to the number of its junction-points, δN_⌓≈((δr)⁄(2)) dN⁄dr where N = (r⁄L_χ )^χ while χ and L_χ are empirical constants.$

3D convolutional networks in deep learning

In medical science, MRI imagery is routinely analysed with 3D convolutional neural networks (CNNs). They identify a certain group of points as belonging to a target entity, such as a tumor (see diagram below).

Any CNN involves neurons organized in layers. Each neuron accepts multiple inputs from neurons in a previous layer, which if super-critical trigger a single output to a neuron in the next layer. When trained, the neural network learns the representation of the data and can perform complex tasks.

Schematic picture of how CNNs are typically applied to MRI imagery. First, measurements are made and there is reconstruction of an image from the raw data in wave-number-space (left panels). This is followed by image restoration to remove noise, and finally registration of imagery classified in various types (right panels). From Lundervold and Lundervold (2018).

Such deep learning technology is being applied in the project to replace the expensive branching scheme. The algorithm will be trained on explicitly computed lightning flashes from AC, resulting in a great speedup of the model.

Aircraft measurements

Two observed cases will be simulated:

MC3E (funded by DoE/NASA): precipitating deep convective systems were sampled by aircraft during May 2011 over Oklahoma, and lightning was measured with a Lightning Mapping Array;

· GO-AMAZON (funded by DoE): Very warm based convective systems in Brazil were sampled by aircraft (G-1, see photo below), and aerosol conditions outside clouds were measured; the occurrence and properties of lightning were measured with the Lightning detection Network (LINET), as well as the GLD360 and STARNET networks.

These measurements of microphysical, dynamical and electrical properties of real storms allow detailed comparison of the simulation with coincident observations. This establishes confidence in the model, prior to its use in analysis of how the storm electrification is controlled by environmental factors later in the project.

The NOAA research vessel Ronald H. Brown, above, played host to the AMF2 for ACAPEX. The G-1 returned to McClellan Airfield in Sacramento, California, for flights over the Sierra Nevada and Pacific Ocean.

The G-1 aircraft of US Department of Energy (DoE) (upper panel) that flew in the Go-AMAZON campaign.

Current progress in project

A postdoctoral scholar was hired in Autumn 2022. In winter 2022, the AI algorithm for representing branching was selected, namely a 3D CNN designed for analysis of MRI imagery. As a first step, it is currently being applied to the training data of a few flashes from AC.

Bibliography

Barthe, C., Molinie, G., and J.-P. Pinty, 2005: Description and first results of an explicit electrical scheme in a 3D cloud resolving model. Atmos. Res., 76, 95-113

Gautam, M., Waman, D., Patade, S., Deshmukh, A., Phillips, V. T. J., Jackowicz-Korczynski, M., Smith, P., and A. Bansemer: “Fragmentation in Graupel-Snow Collisions: New Formulation from Field Observations”. Nature Communications, in review (2023).

Lundervold, A. S., and A. Lundervold, 2019: An overview of deep learning in medical imaging focusing on MRI. Z. Med. Phys., 29, I02-127

Phillips, V. T. J., Formenton, M., Kanawade, V., Karlsson, L., Patade, S., Sun, J., Barthe, C., Pinty, J.-P., Detwiler, A., Lyu, W., Mansell, E. R., and S. Tessendorf, 2020: Multiple environmental influences on the lightning of cold-based continental convection. Part I: description and validation of model. J. Atmos. Sci., 77, 3999-4024

Phillips, V. T. J., and S. Patade, 2022: Multiple environmental influences on the lightning of cold-based continental cumulonimbus clouds. Part II: sensitivity tests for its charge structure and land-ocean contrast. J. Atmos. Sci., 79, 263–300

Waman, D., Patade, S., Jadav, A., Deshmukh, A., Phillips, V. T. J., Gupta, A. K., and A. Bansemer, 2022: Dependencies of four mechanisms of secondary ice production on cloud top temperature in a continental convective storm. J. Atmos. Sci., 79, 3375–3404