AI’s Promise in Weather Forecasting: A Case Study

Matthew Kearney

Professor John Horgan

Seminar in Science Writing: HST 401

September 7, 2024

AI’s Promise in Weather Forecasting: A Case Study

Machine Learning Weather Prediction (MLWP) may soon bring significant developments to current meteorology. Numerical Weather Prediction (NWP) has been our historical weather forecasting agent. Many recent research breakthroughs show great applications for MLWP, with similar or higher accuracy levels in a percentage of the time compared to modern systems. As for any technology, we should consider whether we need it in the first place. Is MLWP a rabbit hole of more convoluted, detail-filling science, or is it revolutionary? What can we take from it?

Stephen Hawking gave a solid framework for modeling reality in his 1980 inaugural lecture at the University of Cambridge, arguing that scientific progress may soon disclose a unified theory that would “describe all possible interactions.”. He addresses two requirements for modeling: (1) local laws ‘obeyed’ on a structural hierarchy by various physical quantities and (2) boundary conditions that inform the state of some regions through time, propagating reality (Hawking 1). This can be understood as the ‘laws’ – the forces and interaction of nature (usually expressed in differential equations), and the state – the matter and particles we see. Could MLWP systems learn and improve our theoretical knowledge of physics and bridge the gap toward a unified theory of theoretical physics? Are there ‘laws’ of nature, or are laws convenient?

Machine learning methods have become increasingly integral to our current weather forecasting. Downscaling supports finer-resolution predictions from global weather models for regional weather forecasting. Bias correction improves model performance using a dynamic weighting system to compare the quality of predictions, ultimately giving better predictors more power in the model. But this seems to be just the beginning of a fruitful relationship.

Researchers behind the November 2023 Google DeepMind study Learning skillful medium-range global weather forecasting developed a real-time global weather forecasting model using a graph neural network called GraphCast, trained on almost 40 years of historical global climate data, that proves overall more accurate and efficient than modern meteorological products in NWP. The study also concludes that further MLWP should be researched for specified tasks including but not limited to extreme temperature, atmospheric rivers, and tropical cyclone tracks. ECMWF has the experiential operational model available now!

GraphCast makes record advancements in the current state of weather prediction, generating global forecasting in less than one minute, compared to a scale of hours in modern forecast systems. Instead of relying on solving numerical methods with high-performance computers, GraphCast learns from weather data, capturing crucial information about nonlinear structures in the weather patterns, to generalize accurate forecasts in real time.

GraphCast also performs research on ‘recency’, appending subsequent years of climate data to train multiple models. This approach helps the model generalize to new, unseen data, such as that of our evolving industrialized climate. The researchers fairly “speculate the recency effect allows recent weather trends to be captured to improve accuracy,” (DeepMind), which certainly motivates the further study and danger of climate change. However, the study may have overlooked integral issues of confirmation and sampling bias. 

The researchers quickly assume all improvement comes from the model capturing more recent weather trends, however do not properly consider the possibility of any improvement stemming from the increased volumes of training data in subsequent models which would certainly improve model accuracy as well. More rigorous bias mitigation knowledge should be pursued that would eliminate the introduction of bias to properly attribute climate change.

This study could make the training data for subsequent recency-updated GraphCast of the same data/time size instead, across the same time range to more accurately attribute improvements in performance to 1) just having more data or 2) it truthfully being the effect of capturing climate change in the model.


More research is necessary to validate the capacity of MWLP further for its deep insight into our natural world, and as with any other technology, be evaluated consistently as to how and where it serves humankind. The new computational paradigm of AI has arrived at the front door of scientific and human pursuit and is hitting the ground running. An essential question before us considers our boundaries with this advancement. 

Automation and unemployment, social engineering, and capital-driven authoritarianism may be considered (by some) as increasing extremities on the dangers stemming from artificial intelligence, but are more actualized and one-in-the-same as every day passes. What ideals are we moving towards?

Our communities must be able to identify signs of discordance with humankind and envision more creative forms of scientific production. There is a growing tide of positive directions that our world is moving toward, and we must invite the same social discourse to critique how innovation and our world should be. 













Works Cited

Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton-Rosen, Z., Hu, W., Merose, A., Hoyer, S., Holland, G., Vinyals, O., Stott, J., Pritzel, A., Mohamed, S., & Battaglia, P. (2023). GraphCast: Learning skillful medium-range global weather forecasting. American Association for the Advancement of Science. science.org/doi/10.1126/science.adi2336. 382(1), 6677

Horgan, J. (2024). Huge study confirms science ending! (Sort of). John Horgan (The Science Writer). johnhorgan.org/cross-check/yrb9e7uefpeqrlkiasoc6octxtnm5g.

Hawking, S. W. (1981).  Is the End in Sight for Theoretical Physics?. Physics Bulletin, 32(1), 15

Comments

Post a Comment

Popular posts from this blog

Scaling the Potential of Vertical Farming Going into 2025 and Beyond

Image
  Matthew P Kearney Scaling the Potential of Vertical Farming Going into 2025 and Beyond      Vertical farming may seem like something out of a futuristic movie, but they are very real in today’s agricultural industry. Vertical farming may promote a stronger relationship between urbanization and sustainable crop production. Based on recent innovations across agriculture and technology, vertical farming has the potential to revolutionize our current farming system. Through the development of vertical farming-based production lines, cities and otherwise unfarmable areas can be provided with fresh, local, and pesticide-free produce. Vertical farming and other adaptations of agriculture are not new ideas, but instead a historic and human process.      These technologies are ready to meaningfully address global food security through the technological and computational advancements in agriculture. Vertical farming has a promising future, which becomes a case...
Read more

Knot Your Average Problem: How do Tongue Ties Impact Oral Myofunctional Health?

Image
  Hadia Hussain      Knot Your Average Problem: How do Tongue Ties Impact Oral Myofunctional Health? Sitting in the nursery of the dental office where I worked, I reviewed Baby V’s medical history and reports from the Emergency Department at Jersey City Medical Center. I found it odd that, despite multiple visits and clear signs of malnutrition, no physician had been able to diagnose the cause. Baby V, at a low weight of 12 pounds was struggling with difficulty nursing and bottle feeding, excessive gassiness, colic, clicking sounds when feeding, milk dribbling out of the mouth, and mouth breathing—but doctors had dismissed these signs, labeling him as "healthy."  It didn’t make sense as his low weight raised serious concerns.  According to Pediatric specialists Jenny Olson, DNP, and Sarha DeVrieze, ARNP, at the UnityPoint Clinic, a 9-month-old baby boy should weigh anywhere in between the range of 17 to 25 pounds. Baby V's situation was clearly a ...
Read more

Crisis to Care: NJ’s Battle with Addiction and Homelessness

Image
Vrinda Modi Crisis to Care: NJ’s Battle with Addiction and Homelessness      “Roll a rig!”, I hear over the radio a second after the tones drop. As an EMT in Hoboken, it's all too common for me to respond to at least one 911 call per shift for an individual under the influence passed out on the sidewalk—usually flagged by a concerned bystander or a police officer who believes they might need medical attention. More often than not, the real issue isn't just medical; it's societal. These individuals, often homeless, have become part of the city’s landscape, invisible to most but impossible to ignore. When I ask for their address, it's almost always the local shelter. When we arrive at the ER, we’re often greeted by familiar eye-rolls from disgruntled hospital staff. The ER clerk barely glances up, lets out a sigh, and asks, “Substance abuse?” I nod. “Who is it this time?” By now, the hospital staff know each chronic substance abuser in the city by name. Behind the hospit...
Read more