New Thoughts on the Human-Machine Mix in Weather Forecasting

With the development of digital computers in the 1940s, the stage was set for numerical weather prediction models based on the equations governing the atmosphere, as envisioned by such meteorological pioneers as Andrie S. Monin, Vilhelm Bjerknes, and Lewis Fry Richardson.  Numerical solution of those otherwise unsolvable equations was the catalyst for a revolution in the science of meteorology, and a continuing debate about the role of humans in weather forecasting.  Sverre Petterssen and Werner Schwerdtfeger, among others, began to anticipate how computer forecasts could compete with humans in the task of weather forecasting.  With the introduction of post-processing methods for turning the gridded variables of a numerical model into actual weather forecasts, Leonard Snellman recognized what he saw as a very real possibility:  fully automated public weather forecasting.  Snellman coined the term meteorological cancer to describe the eventual demise of human intervention in the forecast process.

The notion of the human-machine "mix" has been around since at least the 1970s.  The model developers and those using models as input for objective weather forecasting schemes have steadfastly denied their goal is to replace humans in the forecast process.  As I see it, anyone working to develop objective "guidance" for forecasters is basically in the business of replacing humans with their product, whether they admit it or not - or whether or not they even realize that's what a very successful "guidance" product will do.  As model forecasts improve - which they have done continuously since they began - the need for humans diminishes.  For "ordinary" weather situations, it can be argued that humans already no longer add value to the forecast, even at relatively short range.

The use of numerical models has evolved considerably over those first tentative steps at numerical weather prediction.  The models moved rapidly away from crude one-layer models with coarse resolution and very limited physical processes, to today's models based on the so-called "primitive equations" using vastly increased time and space resolution, fully 3-dimensional, and with extensive physical parameterizations, coupled with sophisticated post-processing schemes to convert gridded variables to sensible weather, and even text generation for fully automated forecasting.  The role of humans during this process has been one of "gap-filling" - the limitations of numerical models represented gaps where a human forecaster could add value to the automated products.  With time, the gaps continue to be filled as the technology of numerical weather prediction evolves.  There are fewer and fewer niches where humans have much of a chance to add value.  The gaps are disappearing.

I've talked about this before, in many essays that can be found here.  Recently, it came to my attention that something interesting is being explored in the UK, whereby forecasters could work with models interactively.  Up to now, computer-based forecasts were like the pronouncements of an oracle, and forecasters were faced with either accepting what the models said or rejecting that solution and providing their own alternative forecast by whatever means they had at their disposal.  Forecasters have been similar to high priests in the business of interpreting oracular pronouncements.  This has not been a truly interactive human-machine relationship. 

What I've envisioned for an interactive relationship is that the forecaster would use the model as a tool to test various possible scenarios in a dynamically consistent way.  What if the moisture available was actually greater than the initial conditions for the model showed?  What if the trough approaching was stronger or approaching more slowly?  How would the forecast change?  A forecaster educated and trained properly could use the model to test such possibilities intelligently and efficiently, and to see the ramifications of those "what if" scenarios.

As I now see things, if something of this sort is not explored and developed, virtually everything now done by forecasters eventually will be automated.  The only debate will be how soon full automation will take place.  Meteorological science is spending a considerable effort all time trying to improve the model guidance, by whatever means necessary.  What are we doing to refine the role of humans and to improve their performance?  Damned little!! Remember:  highly accurate guidance = no more need for forecasters!  Humans cost much more than computers. 

An interactive relationship between model and forecaster would demand a considerably more comprehensive grasp of the science by the forecaster than is now the case.  And it would require a much more extensive training program for human forecasters.  Today's forecasters need to consider their future - young entry-level forecasters may find themselves out of a job before they're old enough to retire!  No one in public weather forecasting is safe from this.  NO ONE!!