climate-weather/ai produced the result/Royal Society Open Science 2022 · v2
Neural network trained on simple equations spots tipping points in real data
Researchers simulated time series from nine ecological, biological and palaeoclimatic models, then trained a neural network to sort them into catastrophic, non-catastrophic or no transition. The trained network was then applied, without retraining, to published real-world records.
spectrum · one line per step, placed by what the step does · bright lines used AI
Machine learning methods trained on simple models can predict critical transitions in complex natural systems
Royal Society Open Science, 2022
doi:10.1098/rsos.211475 · record aix-00294 v2 · checked 2026-11-05
- AI was for
- Classification, Forecasting
- Model family
- Recurrent neural network, Convolutional neural network, Random forest, Support vector machine, Multilayer perceptron, Linear model
- Checked by
- Held-out13 tested, 9 worked
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Many natural systems can flip suddenly from one state to another. A lake turns cloudy, a fish population collapses, a climate regime shifts. These are often called critical transitions, or tipping points, and some are catastrophic, meaning the system does not simply slide back when conditions are reversed. Scientists have long looked for warning signs in the wobble of measurements taken before a flip, using general statistics such as how much a signal varies and how strongly each measurement resembles the one just before it. The difficulty is that real records are short, noisy and messy, and the same statistic can behave differently from one system to the next.
The researchers set out to see whether a network trained entirely on data from simple mathematical models could then classify transitions in real measurements it had never seen. They also wanted to know how it compared with classifiers built on the conventional warning statistics.
Where AI came in
The AI here is EWSNet, a neural network built from two parts trained together from scratch: a long short-term memory branch, a design that reads a sequence step by step and carries forward what it has seen, and a fully convolutional branch, which scans for local patterns in the signal. The team generated noisy time series from nine model systems under two kinds of noise, labelled each series by transition type, and trained one network per noise regime on 80 per cent of the data, feeding in the raw series rather than hand-computed statistics.
The network therefore stood in for the conventional step of choosing and interpreting warning indicators by hand. For comparison, the authors computed those generic indicators over various window settings and fed them to four standard classifiers: logistic regression, a support vector machine, a random forest and a multilayer perceptron. The networks were then tested on withheld simulated data, in a protocol that held out one source model at a time, and on published palaeoclimatic, whaling and laboratory protist records.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The authors trained a deep neural network (EWSNet), combining an LSTM branch and a fully convolutional branch, on stochastic time series simulated from nine ecological, biological and palaeoclimatic models labelled as showing a catastrophic transition, a non-catastrophic transition, or no transition. On the withheld 20% of simulated data, mean test accuracy was 99.46% for the white-noise dataset and 95.93% for the coloured-noise dataset, higher than four classical machine learning classifiers trained on generic early warning statistics such as standard deviation and lag-1 autocorrelation. Applied without retraining to published palaeoclimatic, whaling and laboratory protist time series, EWSNet assigned the expected transition class in 9 of the 13 series whose label was known in advance; in a leave-one-model-out test, 6 of 8 trained variants predicted transitions in data from the held-out model.
How AI was used
Time series were simulated from nine dynamical models under white and coloured noise using the Euler–Maruyama method, and labelled by transition type. Two instances of EWSNet (three stacked convolution blocks with batch normalisation and ReLU, plus a 128-unit LSTM branch, concatenated into two fully connected layers, with global average pooling to make the model invariant to sequence length) were trained from scratch on raw univariate series, one per noise regime, using 80% of the generated data, with hyperparameters tuned on training and validation splits. For comparison, generic early warning indicators were computed over combinations of bandwidths and rolling window sizes and used as input features to logistic regression, a support vector machine, a random forest and a multilayer perceptron, all fitted by the authors. The trained networks were then run over the held-out simulated split, over re-sampled series, in a leave-one-model-out protocol, and over published real-world and experimental time series of varying length, with results averaged over 25 trials.
The shape of the work
Structural · the record, drawn
Simulate training time series from dynamical models
Numerical or physics simulation, including where a learned surrogate replaces it.
We have generated stochastic time series from nine different models ranging from ecological to palaeoclimatic systemswhere the paper describes this · verbatim
Train EWSNet on raw simulated series
Fitting model parameters, including fine-tuning an existing model. The AI stood in for conventional algorithm.
Two different EWSNet models were each trained independently on Dataset-W and Dataset-C, respectively, using 80% of the generated dataset.where the paper describes this · verbatim
Compute generic early warning indicators
Encoding data into features, descriptors, embeddings or graphs.
Generic EWSs are calculated using time series for a combination of bandwidths ({20, 30, 40}) and window sizes ({40, 50, 60}).where the paper describes this · verbatim
Train classical ML baselines on indicator features
Fitting model parameters, including fine-tuning an existing model. The AI stood in for conventional algorithm.
We use classical ML models, namely logistic regression, random forest, SVM, MLP, to classify time series based on the extracted EWSswhere the paper describes this · verbatim
Evaluate on held-out simulated test split
Testing outputs against ground truth.
The remaining 20% was used for testing the performance of the trained model.where the paper describes this · verbatim
Leave-one-model-out cross validation
Testing outputs against ground truth.
The EWSNet is trained using data from eight models each time and tested on the data from the left-out model.where the paper describes this · verbatim
Classify real-world and experimental time series
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
we tested it on 17 published palaeoclimatic and ecological datasetswhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The paper's result is the trained network itself and the transition classifications it produces for simulated and real-world time series
EWSNet is a parametrized function deploying long short-term memory (LSTM) and fully convoluted network (FCN) sub-modules.where the paper describes this · verbatim
The EWSNet classifies transitions in 9 out of 13 real-world time series (for which the ground truth label is known a priori)where the paper describes this · verbatim
Codes and data are available in a Zenodo repositorywhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of EWSNet (LSTM + fully convolutional network)Which version of the model was used is not stated.
- Version of logistic regressionWhich version of the model was used is not stated.
- Version of support vector machineWhich version of the model was used is not stated.
- Version of random forestWhich version of the model was used is not stated.
- Version of multilayer perceptronWhich version of the model was used is not stated.
- What step 5 replacedThe paper gives no basis for what the AI stood in for.
- What step 6 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00294, version 2, checked by a person on 2026-11-05. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error
Cite this record: AIxSci (2026). Neural network trained on simple equations spots tipping points in real data. Record aix-00294, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00294/