One Antarctic glacier could soon lose mass at the same rate as the entire continent

If Thwaites continues along its most concerning trajectory, it could be losing 180–200 billion tonnes of ice per year by 2067 – comparable to the total mass loss from all of Antarctica today. This is why Thwaites has long been a focal point for glaciologists: its mass loss has increased sharply since the 1990s, and its geometry makes retreat difficult to stop once it begins.

In the new study published in AGU Advancing Earth and Space Sciences, computational glaciologist Daniel Goldberg and colleagues at the School of GeoSciences, University of Edinburgh, United Kingdom and the Department of Earth Sciences at Dartmouth College, Hanover, NH, United States ask not just how much ice Thwaites may lose but a more fundamental question: which observations can we trust when predicting its future?

“Thwaites is an important glacier to look at, but the larger context is that what we did in this article is almost a proof of concept. The glacier is just the example. The deeper question is how you build models that point to the right result – and with Thwaites, model uncertainty outweighs all other factors for the next century.”

The study therefore shifts from focusing on a glacier in rapid change to a broader question about prediction itself – and how easily it can be steered. As Goldberg puts it, “the method of calibration is not a technical footnote – it changes the future you get.”

Projections of sea-level rise depend as much on how models are trained as on the ice itself.

A glacier already slipping out of balance

The Thwaites Glacier in West Antarctica has become one of the clearest examples of how quickly a seemingly stable ice system can shift out of balance. Over the satellite era, total ice loss from Antarctica has increased substantially, and a large part of that change comes from glaciers in the Amundsen Sea sector. Among them, Thwaites stands out – both for what has already happened and for what may come next.

The glacier’s ice discharge has increased sharply; its mass loss has grown more than fivefold since the 1990s, and the grounding line has retreated at an accelerating pace. For researchers, this is a warning signal: Thwaites rests on a bed that slopes inward and deepens inland. As the grounding line retreats into deeper water, thicker ice becomes exposed to the ocean, making the retreat self-reinforcing.

“Thwaites is a good glacier for testing this because it is such a strongly dynamic one. In recent intercomparisons, model uncertainty for Thwaites dominated everything else for the next century. That means there is big disagreement in models when modelling this glacier – so it was a difficult one but a very good one to test.”

Concerns about Thwaites are not new. Previous modelling studies have suggested that the glacier could be heading towards an unstable future. But uncertainty has remained high – especially regarding timing. Is this a process that will only accelerate over centuries, or is a major shift already underway in this century?

That question has become more pressing as satellite observations have improved – and enabled models to be tested more directly. Researchers can now track how fast the ice is moving, how the surface is lowering and where melting is strongest – so that models can be tested far more rigorously than before.

“A very small subset of ice-sheet modellers actually try to assimilate the time-dependent aspect of satellite data. That is possible to do now because there are more data – with a longer history. Since Mathieu Morlighem and I both had models capable of doing this, we thought it was time to do a proper comparison between the two.”

This set the stage for the study’s main question: Thwaites is both one of the most concerning glaciers on Earth and the kind of demanding test case that could reveal whether better methods change the answer. As Goldberg explains, model uncertainty for Thwaites “dominates everything else for the next century,” making it exactly the kind of glacier for which better methods might matter most.

How scientists train the model changes the future

To understand how Thwaites may evolve towards the middle of this century, Goldberg and colleagues focused on one of the most critical – and often overlooked – steps in climate modelling: how models are trained on real-world data – by adjusting internal parameters until the simulated glacier acts like the observed one.

For Goldberg, that was the main point of the article. Thwaites served as a proving ground – “the glacier is just the example” – on how to build models that more reliably capture how ice sheets evolve.

The researchers tested this using two independent ice-sheet models constrained by satellite data from 2004 to 2017, run under three calibration strategies: velocity, changes in surface elevation and both combined,

This may sound technical, but the principle is simple: models must reproduce the present to say something reliable about the future – and the key question is which observations best capture the processes controlling how the ice flows and thins. For Goldberg, that decision is as important as the calibration method itself.

“There is a view that once you have this way of adjusting parameters in order to get something to agree with the time-dependent trajectory, everything will work out,” he says. “But I think that is half – or less than half – of the work. You still need to decide what observations you want the model to fit.”

Seeing the glacier over time changes what the future looks like

In this study, that choice changed the future the models produced, using time-dependent calibration to match not just where Thwaites is but how it has changed over time – capturing its trajectory rather than a single snapshot.

Goldberg compares the difference to hanging a picture on a wall: “If you are hanging a picture on a wall and you are standing close enough to touch it, you cannot step back far enough to see whether it is actually level. It may look straight from close up, but then you step back and it is not level. What these methods let you do is adjust the picture with arms long enough to see from across the room whether you are doing it correctly.”

That is what time-dependent calibration makes possible: not just matching where the glacier is but also whether the path reflects the underlying physics of how ice actually flows.

Once calibrated using the various approaches, the models are then run forward to 2067. All other factors are held constant, so that whatever differences emerge can be traced back to one choice alone: the data used to calibrate them.

The same glacier can produce very different futures

The result is effectively a controlled experiment: the same glacier and physics, but different ways of learning from reality. This remains rare in ice-sheet science, in part because it is technically demanding.

As Goldberg explains, once time dependence is included, researchers adjust “thousands, tens of thousands, hundreds of thousands” of parameters, requiring more advanced optimisation methods.

“We are adjusting thousands, tens of thousands, hundreds of thousands of parameters. Traditional parameter-estimation methods will not work. The closest popular analogy is probably neural networks: they also have huge numbers of parameters that are adjusted iteratively to reduce error. We do something very similar.”

Goldberg’s and Morlighem’s models use a technique called automatic differentiation to speed up this iterative adjustment – an approach similar to how neural networks are trained.

The method enables them to ask a deeper question: not just what will happen to Thwaites but how their choice of data shapes the answer. Goldberg describes the study as “almost a proof of concept”: a first demonstration that this way of calibrating models can change what they agree on and therefore what they predict. Nevertheless, he is careful not to overstate the result.

“I do not want it to sound like we are doing the right thing and others are doing the wrong thing. I do not think that is fair. We did the best that we could.”

The most alarming scenarios appear when the model fits the surface

When the models are run forward to 2067, a clear pattern emerges: the future of Thwaites depends strongly on how the model is trained – because different data emphasise different processes inside the glacier.

All models reproduce the present – but not with the same accuracy. Models calibrated using surface elevation changes reduce the error in ice loss by up to an order of magnitude compared with other approaches. Just as importantly, they bring two independent models much closer to the same answer. This matters because Goldberg is not inclined to trust model output blindly.

“I consider myself to be a sceptic when it comes to any model predictions of climate change,” he says. “I would not have believed our results at all if one of our models had done something completely different than the other.” In that context, agreement between independent models becomes a meaningful part of the result rather than a technical detail.

Projecting forward, the models diverge sharply. Those calibrated using velocity data alone show relatively high losses at first but then stabilise. Models calibrated using surface elevation changes do not stabilise; their losses continue to accelerate, reaching 180–200 billion tonnes per year by 2067.

The biggest difference lies deep inside the glacier

What stands out is that the divergence arises not from the physics but from the data used to calibrate it. Goldberg suspects that when velocity data are emphasised, models may end up “overfitting to the velocities” and miss changes that are more directly tied to retreat. More data do not automatically improve a forecast – if they emphasise the wrong part of the system.

“My suspicion is that if you use a specific type of observation, such as velocity, and the coverage is not quite right or the errors are not quite right, you might overfit to those velocities at the expense of fitting correctly to surface elevation change. I think that is what happened to us.”

Looking at where the ice is thinning reveals the next layer of that difference. In the most concerning scenarios, thinning spreads up to around 100 kilometres inland along a deep trough beneath the glacier – bringing faster flow into thicker ice and reinforcing further loss.

“In the surface-constrained runs, thinning penetrates deep inland along the trough – affecting thicker ice further upstream.”

Where the glacier starts giving way matters most

In the most accurate models, the acceleration is not evenly distributed but concentrated in narrow zones along the deep trough – unlike models based on velocity data, in which the changes appear more diffuse.

“The spatial pattern is the key – the realistic models show focused inland acceleration and not a diffuse response.”

To reproduce the observed changes, the models must reduce resistance to ice flow along the deep trough – enabling ice to accelerate and transmit that acceleration inland.

“To reproduce the observed thinning, the model has to weaken the bed along the trough – and that is what unlocks the inland response. In the surface-constrained runs, thinning penetrates deep inland along the trough. That inland signal is critical, because it means you are affecting thicker ice further upstream. It is not just more loss – it is a different spatial pattern of change.”

These same regions of accelerated flow largely control future ice loss, because they connect fast-moving coastal ice with the thicker interior – and transmit change inland. That matters because, as Goldberg sees it, the point is not just to make models more sophisticated, but to identify which signals and which parts of the glacier actually matter most for what happens next.

“The regions that control future loss are the same regions where we see inland acceleration today.”

The clearest warning may be written on the surface

A key finding is that models using surface elevation data reproduce the observed evolution more accurately – even without velocity data. That was not the outcome Goldberg expected reviewers to focus on, but it became central to the article because the results seemed to get worse, not better, when velocity observations were added.

“The biggest critique from the reviewers was not that we generally did the wrong thing but that things seemed to get worse when we included velocities in our observational dataset. I think that is because we ended up overfitting to the velocities at the expense of seeing important things that were more tied to retreat.”

The reason is that surface elevation reflects the combined effect of many processes – from basal friction to melting and ice flow – and therefore acts as an integrated measure of the glacier’s overall state.

“Surface elevation change is an integrated signal of everything happening in the glacier – velocity is just one part of that picture. It is telling us something fundamental about what matters most in the observations.”

The models may agree on the present and still diverge on the future

These differences are not just temporary. If one looks only at total ice loss, the models may appear similar – but the processes driving that loss, and how it evolves over time, diverge sharply: some stabilise, whereas others accelerate rapidly.

“If you only look at total loss, you might think the models agree – but the rate of change tells a completely different story.”

By the end of the period, ice loss in the most extreme scenarios is more than twice as high as in the other models – and comparable to the current total mass loss of the Antarctic Ice Sheet.

“By 2067, the loss rates in these models are comparable to the current mass balance of the entire Antarctic Ice Sheet. That puts one glacier into a global context in a very direct way.”

The results point to a double realisation: Thwaites is already on a trajectory towards increased ice loss, and the assessment of how fast that loss may grow depends strongly on which data we trust – and therefore which processes we enable the models to give priority to.

We may be underestimating how fast the loss can grow – and why

If a single glacier reaches an annual ice loss of up to 200 billion tonnes, the consequences are no longer just Antarctic – they are global. Goldberg is careful about how that risk is described. Terms like “doomsday glacier”, he suggests, may cause people to switch off rather than listen. The real need is for more reliable estimates of what is likely to happen in Antarctica in the near future – and for willingness to say what the models show even when the message is uncomfortable.

“We need to not be afraid of messaging that might be counter to what people would want to see,” he says.

Thwaites alone holds enough ice to raise sea level by more than half a metre if it were to collapse completely. But the new study points to something just as important: uncertainty is not only a question of physics – it is also a question of method.

“Taken together, these results are difficult to ignore. If the choice of calibration data determine whether models show stabilisation or continued acceleration, then some projections of future sea-level rise may be underestimating – or overestimating – the risk.”

When different models tell different stories

According to Daniel Goldberg, the challenge is not only scientific but also communicative: different models produce different answers – and, with them, different stories about the future.

As he puts it, this creates “a recipe for many, many different possibilities depending on which model is used,” and therefore “lots and lots of different messages.” That, in turn, makes it easier to pick and choose the conclusion one prefers. “There should not be such uncertainty,” he says. “And if there is any way to address that uncertainty, I think that should be jumped upon.”

This is not the last word, however. The study covers only the next 50 years or so and does not include all processes that may influence the glacier over longer time scales – such as evolving ice structure, subglacial water flow or the potential collapse of floating ice shelves that currently slow the glacier.

Goldberg is explicit that this is not a final answer but an initial demonstration of what this modelling strategy can do. “That is why this is a proof of concept,” he says. “The article represents the best way we could think of, with the experiments we did, to evaluate these questions – but there is definitely room to expand.”

The study is best understood as a carefully argued starting point rather than a definitive conclusion.

What the glacier is already telling us

The study does not identify a single correct way to calibrate models but highlights the need to use all relevant observations – and to understand which physical processes each actually captures. Today, velocity data are often given significant weight, but these results suggest that changes in surface elevation may contain crucial information about the deeper dynamics of the ice.

“My hope is that the next generation of models will look at this study and studies like it and think about a better way to do things.”

In that sense, the study extends beyond Thwaites itself. It becomes an example of a broader challenge in climate science: our ability to predict the future depends not only on better models but on how well we read the signals already written into the present.

Advances in modelling – partly driven by techniques from artificial intelligence – mean that ice-sheet models now under development are better equipped to assimilate time-dependent observations, similarly to the models used in this study. Goldberg hopes the lessons learned will carry into the next generation of ice-sheet models. Even if today’s models are replaced, he wants future ones to take from this work “a better way to do things.”

Goldberg is also careful about tone. “I am not saying we should not be worried,” he says. “But perhaps terms like ‘doomsday glacier’ are why people just switch off entirely. I think we need more careful wording.”

His aim is not to simplify the science, but to strengthen the modelling – and make the message harder to dismiss.