Reflections on 25 Years of PEST Courses

A quarter-century of teaching groundwater scientists how to extract decision-relevant information from numerical models has prompted one of the field’s most experienced practitioners to reflect on what a generation of training courses has taught him about calibration, uncertainty, and the often uncomfortable gap between what models can deliver and what decision-makers expect of them. John Doherty, the creator of the PEST software suite and founder of Watermark Numerical Computing in Brisbane, Australia, has published a retrospective essay in Hydrogeology Journal documenting lessons drawn from 25 years of conducting courses on model calibration and uncertainty analysis. The essay, which appears as the field grapples with ever-larger environmental questions and ever-more complex simulators, argues that education in these technical disciplines does far more than transfer skills: it shapes how the entire groundwater modelling industry understands, implements, and evaluates decision-support modelling.

Doherty’s central contention is that calibration and uncertainty analysis are not merely technical chores appended to the modelling workflow. Instead, they lie at the heart of what groundwater modelling can, and cannot, contribute to environmental decision-making. When practitioners learn these subjects properly, they acquire the skillsets needed for what he describes as simulator-based harvesting of decision-pertinent information from site data. This framing is deliberate. It positions the numerical model not as a replica of reality to be perfected, but as an instrument for interrogating noisy, sparse, and indirect field measurements in ways that yield numbers relevant to the decisions at hand. The distinction matters because a model that reproduces historical observations beautifully may still be useless for predicting the consequences of a proposed pumping scheme, a mine dewatering plan, or a contaminant remediation strategy. The information content of the data, not the elegance of the model, determines what can reliably be learned.

The PEST software, which Doherty began developing in the 1990s, has become one of the most widely used tools for model calibration and parameter estimation in the groundwater community, and it is now applied across broader environmental modelling domains as well. PEST automates the process of adjusting model parameters so that simulated outputs match field observations as closely as possible, and its modern versions implement formal methods for quantifying the uncertainty that surrounds model predictions. Over the decades, Doherty and colleagues have delivered courses on these methods to thousands of practitioners in regulatory agencies, consulting firms, and research institutions around the world. The new essay distills the pedagogical and philosophical insights accumulated across that long teaching record into a document that doubles as a critique of industry practice.

Among the recurring themes Doherty identifies is the persistent tendency to equate a good calibration with a trustworthy model. His teaching experience suggests that many modellers still regard a close match between simulated and observed hydraulic heads, flows, or concentrations as the primary certificate of model quality. The modern uncertainty analysis perspective overturns this assumption. Parameter nonuniqueness, in which many different combinations of parameter values reproduce observations equally well, means that a well-calibrated model can nonetheless make wildly divergent predictions under future conditions. Calibration noise, arising from measurement errors, model structural imperfections, and the numerical approximations inherent in any simulator, further limits the inferences that can be drawn from history matching. Courses that confront students with these realities, Doherty argues, inoculate them against a false confidence that can otherwise propagate into decisions worth hundreds of millions of dollars.

The essay also grapples with how uncertainty analysis should be communicated and, more pointedly, how it should be simplified. Rigorous linear and nonlinear methods for propagating parameter uncertainty into predictive uncertainty can be computationally demanding and conceptually intimidating. Doherty’s courses have long emphasized pragmatic approaches that retain mathematical defensibility while remaining accessible to working modellers facing deadlines and limited budgets. The underlying message is that some quantification of predictive uncertainty, however approximate, is far better than none, because it converts an unexamined gamble into a stated risk. This conversion, in his view, is the essential service that groundwater modelling renders to society. A decision-maker who is told that a proposed water allocation will cause a lake to dry with a probability of 20 percent, rather than being told simply that the lake will survive, is equipped to weigh that risk against economic and social considerations.

Education, in Doherty’s telling, engenders a perspective that reverberates through the industry at large. Practitioners trained to think in terms of information content, parameter vagueness, and posterior uncertainty tend to design data-collection programs differently, choosing monitoring locations and measurement types that maximally reduce the uncertainty of the predictions that matter. They tend to document assumptions more transparently and to present results with honest caveats. Over time, as trained practitioners rise into management and regulatory positions, this perspective can shift institutional expectations, raising the bar for what constitutes an acceptable modelling study. The essay implies that the cumulative effect of a quarter-century of courses is cultural as much as technical, gradually redefining professional norms around calibration and uncertainty in groundwater practice.

The reflections arrive at a moment when the pressures on decision-support modelling are intensifying. Groundwater systems face mounting demands from agriculture, urbanization, mining, and climate variability, while regulators increasingly require quantitative evidence that proposed activities will not violate environmental thresholds. Meanwhile, simulators have grown in sophistication, and parallel computing has made ensembles of model runs feasible at scales that would have been unthinkable when PEST was young. Doherty’s essay suggests that this abundance of computational power sharpens rather than resolves the central challenge: more model runs do not create information where none exists in the data. The bottleneck remains the inferential content of sparse field measurements, and the discipline that teaches modellers to respect that bottleneck is the discipline his courses have championed for 25 years.

The essay’s framing of model calibration as an information-theoretic exercise carries practical consequences for how modelling studies should be scoped and priced. If the goal is to harvest decision-pertinent information, then effort should be concentrated on the parameters and processes that most strongly influence the predictions underpinning the decision, rather than distributed uniformly across a model’s full parameter space. Doherty’s courses have taught techniques for identifying these sensitive parameters and for using calibration-constrained uncertainty analysis to examine whether the data can actually pin them down. Where they cannot, the honest response is to report predictive uncertainty rather than to artificially constrain parameters through unjustified regularization. This philosophy, which PEST’s regularization tools were designed to support, treats the model as a vehicle for making assumptions explicit and testable rather than as an oracle whose outputs are taken at face value.

Doherty’s reflections also touch on the communication gap between modellers and their audiences. Decision-makers, communities affected by water management choices, and even many fellow scientists may find the vocabulary of parameter estimation and Bayesian statistics impenetrable. The courses therefore function partly as translation exercises, equipping modellers to explain in plain language why a prediction carries an uncertainty range, what that range means, and how additional data would narrow it. Doherty argues that this communicative competence is inseparable from technical competence. A modeller who cannot convey the epistemic status of a prediction has not finished the job, because the ultimate destination of modelling outputs is a decision made by people who will never read the calibration report.

Beyond the technical curriculum, the essay carries a note of advocacy. Doherty has spent recent years associated with the Groundwater Modelling Decision Support Initiative, an organization devoted to promoting best practice in the use of groundwater models for environmental management, and the essay acknowledges funding from that initiative over the past five years. Through the initiative and the courses alike, the message has been consistent: the value of a groundwater model lies not in its realism but in its capacity, when used rigorously, to convert field data into decision-relevant insight while being candid about the limits of that insight. The essay’s publication in Hydrogeology Journal places this message before the international hydrogeological community at a time when the stakes of groundwater decisions have never been higher, and when the profession’s collective habits, formed in part by decades of PEST training, will shape whether models illuminate those decisions or obscure them.

The full essay, including Doherty’s own reflections on the evolution of his teaching and the software ecosystem that grew around it, is available in Hydrogeology Journal. PEST and all of its associated software remain freely available for download, ensuring that the tools accompanying this quarter-century of pedagogy remain accessible to practitioners and students worldwide.