Weather Forecasting Limitations in the Developing World

Abstract

The first high performance computing (HPC) capability any developing country deploys is weather forecasting. In the developed world, we take for granted widely available, high quality forecast models that give us 48–72 h of excellent quality and up to 10 days of good quality forecasts. These forecasts enable planning to protect life and property. One recent example of the importance of this capability is in Mozambique. Days of torrential rains led to catastrophic flooding and loss of life. Good quality weather forecasting and effective communication could have warned people to either reach higher ground or otherwise evacuate low lying and flood prone areas.

With extensive computational resources, the developed world continuously not only offers models, but also checks forecasts against measurements to find gaps and errors in the models. As these are identified, they are being tweaked continuously improving forecasting models for that country or group that supports the forecasting center.

This persistent need for extensive, high capacity HPC resources to better protect the population is not available to developing countries that lack the resources to even operate a basic forecasting model. For example, on the African continent, the most advanced HPC resource, by far, is the Lengau machine in South Africa installed in 2016. It is now old and far behind the needs of current forecasting models. The rest of Africa has been receiving machines South Africa’s CHPC and CSIR have been accepting as donations from the developed world and redeploying them, generally in much smaller pieces, across the continent. Even with these resources, the forecasting models are too complex to complete their calculations to offer timely forecasts for the areas in which these machines are installed.

The recent developments of Machine Learning (ML) models to replace complex physics in systems where humanity either does not understand the physics sufficiently or the physics is too complex to model with current computational resources is extremely promising for rapid solutions on limited hardware. The development of initial weather forecasting models using ML components have been deployed, but these models, like all forecasting codes, are tightly held property of the creating institutions.

Ethically, what obligations does the developed world have to share these potentially slightly less accurate, but vastly more efficient and computationally compact systems with the developing world? This paper explores this topic from many directions seeking to understand the issue more completely.