Machine learning and the quest for objectivity in climate model parameterization - Research

Aug 2, 2026

Upload by Anda

Parameterization and parameter tuning are central in climate modeling and are widely regarded as involving subjective elements. The paper investigates whether machine learning can replace or reduce this subjectivity by offering more automated and ostensibly objective methods. Based on multiple case studies, it argues that ML can improve aspects of climate model parameterization, yet expert judgment remains necessary, carrying subjective components similar to traditional tuning. Careful supervision is therefore essential.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages19

Reading Time48 min

AccessFree to download

Abstract

Introduction

Climate models and sub-grid parameterizations

Sources of subjectivity and underdetermination

Machine learning as a proposed remedy

Research aim and critical assessment

Machine learning and the quest for objectivity in climate model parameterization Julie Jebeile1,2,3 · Vincent Lam1,2,4 · Mason Majszak1,2 · Tim Räz1 Received: 29 September 2022 / Accepted: 7 April 2023 © The Author(s) 2023 Abstract Parameterization and parameter tuning are central aspects of climate modeling, and thereis widespread consensus that these procedures involve certain subjective elements. Even if the use of these subjective elements is not necessarily epistemically problematic, there is an intuitive appeal for replacing them with more objective(automated)methods, such as machine learning. Relying on several case studies, we argue that, while machine learning techniques may help to improve climate model parameterization in several ways, they still require expert judgment that involves subjective elements not so different from the ones arising in standard parameterization and tuning. The use of machine learning in parameterizations is an art as well as a science and requires careful supervision. Keywords Climate modeling · Parameterizations · Parameter tuning · Objectivity · Subjectivity · Expert judgement · Machine learning · Deep neural networks · Gaussian processes 1 Introduction Machine learning applications in science are receiving more philosophical attention as they raise signiﬁcant epistemic issues, related to interpretability and understanding (see Beisbart B Mason Majszak [mason.majszak@unibe.ch](mason.majszak@unibe.ch) Julie Jebeile [julie.jebeile@unibe.ch](julie.jebeile@unibe.ch) Vincent Lam [vincent.lam@unibe.ch](vincent.lam@unibe.ch) 1 Institute of Philosophy, University of Bern, Länggassstrasse 49a, 3012 Bern, Switzerland 2 Oeschger Center for Climate Change Research, University of Bern, Hochschulstrasse 6, 3012 Bern, Switzerland 3 CNRM UMR 3589, Météo-France/CNRS, Centre National de Recherches Météorologiques, Toulouse, France 4 The University of Queensland, School of Historical and Philosophical Inquiry, 4072 St Lucia QLD, Australia 1 3 and Räz 2022; Räz and Beisbart 2022 as well as references therein) . Some discussions have focused on the climate context more speciﬁcally (e.g. Jebeile et al. 2021; Knüsel and Baumberger2020;Kawamleh2021), as thereis high expectation from thescientiﬁccommunity that machine learning can reduce computational cost and overcome climate model uncertainty. Another important expectation however has been overlooked by philosophers: the hope that machine learning can alleviate the subjectivity of some parts of climate models, in particular the parameterizations and their associated parameter tuning. In this paper, we aim to explicate the sources of subjectivity in climate model parameterization and tuning, and subsequently investigate whether, and how, the diverse available machine learning techniques can make climate models more objective. Models can represent the climate system only up to a certain spatial resolution (grid size), due to numerical constraints and the computational cost of higher resolutions.1 The physical laws, e.g., the Navier–Stokes equations, on which models rely, are thus discretizedona grid tobe implemented on the computer and numerically solved in manageable timeframes. Therefore, processes such as convection (clouds), which typically occur at the sub-grid level and are relevant to climate modeling, need to be represented using so-called parameterizations. Parameterizations are simpliﬁed representations—also qualiﬁed as “mini-models”(Lloyd 2015,61)—of these processes based on phenomenological and theoretical considerations (see also Guillemot 2017 and Winsberg 2018, 47-50) . The phenomenological considerations involve many parameters (hence the term ‘parameterization’), some of which are just consequences of the discretization and parameterization procedures and thus do not correspond to anything in the target system. Parameter tuning is intrinsically part of the building of parameterizations, it is the crucial (and

FAQs

Why does climate model parameterization involve subjectivity?

Parameterizations are simplified representations of sub-grid processes, and the choice of parameter values is not uniquely determined by theory or observations. This underdetermination leaves room for subjective expert decisions during design and tuning.

How is machine learning proposed to improve climate model parameterization?

Machine learning models are expected to reduce computational cost, support uncertainty quantification, and potentially make parameterizations and tuning less subjective through more automated approaches.

Does machine learning make parameterizations strictly more objective?

No. The paper argues that even with machine learning, expert judgment is still required, and the subjective elements are not fundamentally different from those in standard parameterization and tuning. ML use also demands careful supervision.

Aug 2, 2026

Upload by Anda