The Earth is amazing.
Terrescient is a scientist-led, AI-native deep tech company. We build data-driven models of whole Earth evolution with quantified uncertainty, coupled from plate motions to the deep mantle, to power the search for the critical minerals a clean energy future needs.
Discovery is getting slower and more expensive.
Critical minerals drive the energy transition. The deposits that were once easy to find have been found, and each new discovery takes longer and costs more.
Deposits form in ancient environments that no longer exist.
Most ore deposits are hundreds of millions of years old. The tectonic settings that made them have since moved, been deformed, buried or destroyed. AI and ML exploration models evaluate the Earth as it is today. We model the Earth system through time. Plate reconstruction plus mantle structure is the missing history that today's maps cannot show.
The argument in full · Plate tectonic modelling and the energy transition, Earth-Science Reviews, 2022 · [ 05 ]From research programme to product.
A coupled model of the whole Earth through time has long been a research programme, not a product. Running one history took months. Running thousands was simply out of reach. Two recent advances have changed that: ensemble computation is now affordable at the scale the problem demands, and AI has advanced from optimising models to reasoning inside them as they run. That capability has accelerated everything else. What was a career's work is now an experiment.
Machine intelligence, built into the model.
Most AI in exploration sits outside the model, speeding up a workflow or scoring a map. Ours is part of the architecture. Agents act inside the running Earth model, assembling and evaluating the physics, interfacing the surface with the deep mantle, and testing alternative Earth histories against observations. Because they are part of the model, they can do this on their own, at a scale no team of scientists could. Our scientists set the direction. The model reasons as it runs.
Swarms, not monoliths.
Frontier mantle models take days on hundreds of cores chasing one detailed answer at high resolution, and plate reconstructions are still built by hand, one at a time, over years. We run swarms of small, fast models instead. Together they map the solution space that single models only sample.
Expertise sets the direction.
Our scientists define the questions, choose the physics and design the experiments. The agents work within those limits, and every result is checked against the governing equations, established reference models and observations.
Autonomy, with limits.
An agent can assemble, run and score an Earth history without a person in the loop. It cannot change the physics, the data or the question. That line is drawn in the experimental design, and it is visible in every result.
Uncertainty is the product.
Inputs are distributions, outputs are scored ensembles. We never ship one answer when the data support many, and we make the full solution space explorable.
Scale changes the question.
A plate reconstruction in seconds, a mantle simulation in minutes, many orders of magnitude cheaper per history. The question is no longer which history to run, but which histories to explore.
Validated by the literature.
Our models are built on published, peer-reviewed science and are always judged against it. We credit the research our work stands on, and we collaborate closely with the academic groups who produce it.
Every answer can be re-run.
Each result carries the recipe, the data inputs and the model states that produced it, so any ensemble can be reproduced, audited and extended later, by us or by anyone who uses it.
The Terrescient Earth Model. Surface to core, in one loop.
Plate reconstruction, absolute plate motion and mantle convection each stand alone as a model. Coupled, they become more than the sum of their parts: a tectonic reconstruction is tested by the plate motions and mantle structure it predicts, and the deep Earth feeds back into the surface. This loop is where the agents work, running and scoring thousands of alternative histories against observations. Inputs are distributions. Outputs are scored ensembles of whole Earth history.
The agent layer of the Terrescient Earth Model. Orchestrating the model coupling, agents assemble each component from a codex of physical ingredients, test every ingredient against the governing equations and published reference runs, and repair what fails. A deterministic queue dispatches the ensemble and grades each history against observations, pushing findings back to the orchestrator for evaluation. The three scientific model components beneath are deterministic physics: agents work between them, never inside them. Agents propose; model tests decide.
Our probabilistic deep time plate tectonic reconstruction model. Plate models have always been built by hand and judged by eye. Continuum is parameterised: every continental position and plate boundary is a variable, scored against the data rather than drawn by an expert.
In palaeomagnetic data as the foundation. All other data is a constraint.
Out a scored ensemble of plate reconstructions, including boundaries, mapping the solution space the data support.
Our absolute plate motion model, built on optAPM. Continuum's reconstructions are born in the palaeomagnetic frame. Keel moves them into the mantle's, and treats the frame as something to test, not assume: a history that fits the magnetic record must also fit how plates can move over the deep Earth.
In Continuum reconstruction plus geodynamic constraining data.
Out the same reconstruction anchored to the mantle, and constrained at every age, including the ones palaeomagnetism cannot see.
Our mantle convection model, built for ensembles. The mantle keeps a record of where plates have been: slabs sink for hundreds of millions of years and tomography can still see them. Mantlepiece runs the mantle forward under each history and compares what it predicts with what seismology sees, in minutes rather than days.
In the same Continuum reconstruction, now in Keel's mantle reference frame.
Out the mantle each history predicts, scored against tomography, so that a reconstruction and its frame are judged by the deep Earth they imply.
Ask the Earth. Get the answer.
Our product is the Terrescient Earth Model, built and kept current by Mise and served by Terrescient AI. Ask it where a continent was, which way it was moving, and when, and it answers with the uncertainty on each: a spread of positions, a spread of motions, a spread of ages, every one scored. Models of the Earth have never taken questions. This one does. That opens every published result to the next question, every new dataset to the model, powering ideas as fast as we can think of them.
Securely run your own experiments on your own hardware with your own data, or ask ours: every dataset, every history, every test, in one model you can talk to.
Three ways to use itTerrescient AI.
Ask the model. Terrescient AI is the Earth Model's own assistant: parameterise an experiment, run it, query the ensemble and analyse the result, all in plain language. Ask a follow-up and it keeps the thread: change one parameter, run again, and compare against the last answer. It works with the language model you already use, so nothing you say to it reaches us. For those whose data cannot leave the building, the same Terrescient AI runs on your own hardware, against a snapshot of the model and your own experiments, and nothing goes back.
Data and maps.
Nothing has to change in how you work. Every answer is also data: ranked histories, the maps and grids derived from them and the uncertainty behind each one, in the formats you already use. Open a reconstruction in GPlates, a grid in QGIS or a notebook, a 3D mantle field in ParaView. Plot the distributions, compare histories and models, and read the uncertainty on everything. Each result carries the model version, its parameterisation and setup, so it can be cited, compared and run again.
Tools for any AI assistant.
The same capabilities are exposed as tools that any AI assistant can call. Whatever system you already work in, if it has an AI assistant, it can have the Earth Model. Ask from wherever your assistant lives, and the answer comes back with its uncertainty. The same operations are an API, with a Python client, for pipelines with no assistant in them: call the model from a script, a notebook or a scheduled job, pull an ensemble straight into a dataframe, and build on it as you would on any other library.
A frame for anything with a place and an age.
The Terrescient Earth Model is a four-dimensional reference frame. Any dataset with a position and an age can be put back where it formed, carried through time with the plates, placed against the mantle beneath it, and scored. The uncertainty travels with the data: a reconstructed position is a spread, not a point, and every answer says how wide. Critical minerals are the first question we put to it. But we have many more.
Published methods. Proven foundations.
Everything we build stands on published, peer-reviewed research, much of it our own. The Earth Model takes that groundwork and makes it one coupled, testable system.
Not just an increment, a step change.
Selected. Every paper here is co-authored by our team, and each contributes a piece to a model that is far more than their sum. Titles, journals and years only.
- A global plate model including lithospheric deformation along major rifts and orogens since the Triassic. The reference method for deforming plate reconstructions: the plate modelling behind Continuum.
- Extending full-plate tectonic models into deep time. The first billion-year full-plate reconstruction: the approach Continuum extends.
- An updated internet-based Global Paleomagnetic Database. The field's reference palaeomagnetic compilation: the record of position Continuum starts from.
- Deep time spatio-temporal data analysis using pyGPlates with PlateTectonicTools and GPlately. The reference paper for pyGPlates and GPlately, with EarthByte: Continuum's open toolchain.
- Constraining absolute plate motions since the Triassic. The first absolute reference frame derived with geodynamic constraints: the method Keel is built on.
- A tectonic-rules-based mantle reference frame since 1 billion years ago. optAPM applied to the billion-year model and judged by mantle convection: the link Keel and Mantlepiece close into a loop.
- Decoding Earth's plate tectonic history using sparse geochemical data. Sparse geochemical samples discriminate between competing histories: an independent check the Earth Model scores with.
- Geochemical fingerprinting of continental and oceanic basalts: a machine learning approach. Machine learning reads tectonic setting from basalt chemistry: data-driven scoring of histories.
- Assessing plate reconstruction models using plate driving force consistency tests. Velocities must match the forces driving each plate: a physics test that ranks reconstructions without new data.
- Tectonic controls on carbon and serpentinite storage in subducted upper oceanic lithosphere for the past 320 Ma. What subduction has delivered to the mantle over 320 million years: the surface-to-deep-Earth histories the Earth Model explores.
- Subduction and carbonate platform interactions. Carbonate platforms consumed by subduction and the arc CO2 that follows: a plate model constraining Earth-system processes.
- Plate tectonic modelling and the energy transition. Critical mineral deposits formed in ancient tectonic settings that only coupled reconstructions can recover: why Terrescient exists.
Built with the people who built the field.
We work in the open, learning from the research groups that shaped plate reconstruction and geodynamics. Methods are published together, code and data are shared, and our models are tested against the work of the people who know the field best.
Collaborations our work is built on
- The University of SydneyEarthByte, School of Geosciences
- Université Côte d'AzurGéoazur
- GFZ PotsdamLithospheric Dynamics
- Curtin UniversityEarth Dynamics Research Group
- California Institute of TechnologySeismological Laboratory
- UT AustinInstitute for Geophysics
- ENS ParisLaboratoire de Géologie
- University of Helsinki and GTK FinlandPalaeomagnetics Laboratory
What we do, not what we believe.
Our agents run inside a model of the Earth, choosing and testing thousands of histories at scale. That raises a fair question: what keeps them honest? The physics does. An agent cannot change the physics, only propose within it. Every proposal is tested against the governing equations, reference runs and observations before it counts, every result carries its uncertainty, and a scientist can re-run any history and see why it scored as it did.
Scientists set the direction.
Agents assemble and test Earth histories. They do not change the physics, they do not choose the question, and they do not mark their own work: the tests and the scoring sit outside their reach. Every run is reproducible: the same inputs give the same scores.
Every result carries its uncertainty.
We never ship one answer when the data support many. Ensembles are scored, and the spread is part of the product, not a footnote to it.
Every record is traceable.
Every dataset in our platform carries its source and its licence. Our models rest on decades of published research by groups around the world, and we name that work and the people behind it. See sections [ 05 ] and [ 06 ] above.
What flows back.
We build on public data and published methods. Model outputs go back into publications and experiments, and we support the students, researchers and groups who will write them.
Beyond critical minerals.
Critical minerals are where we start, not where the model stops. The same frame holds geothermal and hydrogen, carbon storage, past climate and oceans, the fossil record, even other planets: anything with a place and an age. And where it does guide extraction, the less ground disturbed to learn the same thing, the better.
Acknowledgement of Country
We at Terrescient acknowledge the Ngunnawal, Ngambri-Kamberri and Cammeraygal peoples as the Traditional Custodians of the lands on which we operate and recognise the incredible depth, contributions and sophistication of First Nations knowledges. We pay our respects to Elders past and present.
Backed by programmes. Partnered in research.
The structures behind the work: the programmes we belong to, the research we partner on, and the scientists who advise us.
- Startup programme membershipsAnthropic · Claude for Startups. Anthropic's programme for early-stage companies building on Claude.
- Partner organisationARC Research Hub for Geodynamic Pathways to Critical Resources (GeoPath). An Industrial Transformation Research Hub led by The Australian National University, 2026 to 2031, building a Virtual Earth for Resources: deep Earth dynamics linked to surface processes, with AI used to predict mineral potential from plate reconstructions, mantle and landscape models. Terrescient is an official Partner Organisation. arc.gov.au
- AdvisoryScientific advisory board. We are fortunate to have some of the most senior researchers globally in plate tectonics, geodynamics and mineral systems advising on the science. Their role is twofold: to hold what we build to the standard of the published literature, and guide the scientific direction of our emerging technologies.
One address. A person reads it.
Researchers, explorers, investors, reviewers, people who want to work on the whole Earth: write to us and say who you are. We answer.