Why Science? | Surface Science Foundation Model

Catalyst possibilities multiply across elements, compositions, crystal faces, defects, adsorbates and reaction conditions. Conventional trial-and-error cannot test every combination. The project aims to navigate this space with efficient AI while keeping chemical constraints and validation visible. The challenge is search plus trustworthy evidence, not size alone.

Catalysis happens at surfaces where atoms adsorb, react and leave; tiny changes in local structure can alter activity, selectivity, stability and the energy needed for a chemical pathway. The chapter develops design-space reasoning. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: List five axes of catalyst variation and estimate how combinations multiply. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A large search space does not justify weak validation.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Next chapter

Official sources describe a five-year S$10 million collaboration joining computational materials science, chemistry, AI and automated experimentation. Imperial names co-leads from both institutions and an open cloud-platform ambition. Current roles, outputs and access should be taken from dated project pages.

The terascale challenge comes from the huge number of possible compositions, structures and sites, making efficient representations and carefully chosen experiments central to discovery. The chapter develops programme literacy. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Map computation, synthesis, characterisation and reaction testing to responsible teams. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Funding and collaboration do not prove target outcomes.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

A catalyst offers an alternative reaction pathway and is regenerated through the catalytic cycle, though it can deactivate over time. Useful assessment includes activity, selectivity, stability and operating conditions. Faster conversion is not enough if the process creates unwanted products or consumes excessive energy.

Inverse design begins with desired performance and searches backward for candidates, but synthesis, characterisation and realistic reaction tests must close the loop. The chapter develops catalysis literacy. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Draw a simplified energy profile and identify catalyst performance measures. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A catalyst is not literally unchanged under all operating conditions.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Molecules adsorb at terraces, steps, edges, vacancies or mixed-metal environments. The local electronic and geometric structure affects bond formation and breaking. Models need representations that distinguish meaningful sites without treating every translation or rotation as a new chemistry problem.

Efficient AI matters scientifically because compute, memory and time determine how widely a model can explore and how readily other researchers can reproduce the search. The chapter develops surface-structure reasoning. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Compare three surface sites and predict how coordination might alter binding. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A site hypothesis requires structural and reaction evidence.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

NRF describes training on an estimated two million or more computational and experimental data points. Calculations offer scale and controlled labels; experiments capture real defects, preparation and conditions. The dataset should preserve method and uncertainty so the model does not treat different evidence as interchangeable truth.

Catalysis happens at surfaces where atoms adsorb, react and leave; tiny changes in local structure can alter activity, selectivity, stability and the energy needed for a chemical pathway. The chapter develops scientific data integration. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Create a data card separating computed and measured surface properties. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Dataset size does not remove method bias or inconsistency.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

An equivariant model changes its internal outputs in predictable ways when an atomic structure is rotated, translated or reflected as appropriate. This lets the architecture use physical symmetry rather than relearn every orientation. The chosen symmetry and representation must match the predicted property and be tested on held-out structures.

The terascale challenge comes from the huge number of possible compositions, structures and sites, making efficient representations and carefully chosen experiments central to discovery. The chapter develops geometric-machine-learning literacy. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Illustrate an atomic rotation and which outputs should remain or transform. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Equivariance does not encode every law of chemistry.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

The project includes Equivariant Sparse Structure Learning, combining geometric architectures with sparse techniques. Sparsity can focus computation on informative interactions and reduce resource needs. Researchers should report accuracy, memory, runtime and scaling, comparing against simpler baselines rather than claiming efficiency from architecture names alone.

Inverse design begins with desired performance and searches backward for candidates, but synthesis, characterisation and realistic reaction tests must close the loop. The chapter develops computational-efficiency evaluation. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Build an accuracy–compute comparison with common hardware and datasets. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Lower compute is meaningful only at comparable scientific performance.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Randomly splitting nearly identical structures can inflate results. Strong evaluation may hold out compositions, surface motifs, reactions or time periods and test uncertainty on unfamiliar regions. A model should know when a candidate lies far from its training evidence and route it for careful review.

Efficient AI matters scientifically because compute, memory and time determine how widely a model can explore and how readily other researchers can reproduce the search. The chapter develops generalisation testing. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Design three split strategies and explain what each tests. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: High in-distribution accuracy does not prove discovery capability.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Instead of predicting performance for a supplied material, inverse design searches for structures expected to meet activity, selectivity, stability and cost goals. The objective should include synthesis and safety constraints. Diverse candidates are more useful than many minor variations around one predicted optimum.

Catalysis happens at surfaces where atoms adsorb, react and leave; tiny changes in local structure can alter activity, selectivity, stability and the energy needed for a chemical pathway. The chapter develops inverse-design literacy. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Write a multi-objective target and three hard feasibility constraints. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A generated structure may not be stable or synthesizable.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

NRF says reinforcement learning and Bayesian optimisation will help refine candidates against several criteria. An agent can decide which calculation or experiment to request next. Selection should remain auditable through acquisition values, alternatives, stopping rules and human overrides, especially when objectives conflict.

The terascale challenge comes from the huge number of possible compositions, structures and sites, making efficient representations and carefully chosen experiments central to discovery. The chapter develops agentic optimisation governance. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Create an optimisation decision log with expected gain and realised evidence. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: An agent score is not a chemical conclusion.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Imperial describes integration with automated experimental tools. High-throughput synthesis and testing can produce consistent data quickly when recipes, instruments and reference catalysts are controlled. Negative outcomes should return to the model, while suspected equipment failures remain separately labelled.

Inverse design begins with desired performance and searches backward for candidates, but synthesis, characterisation and realistic reaction tests must close the loop. The chapter develops experimental automation. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Design a robotic screening plate with controls, randomisation and repeat tests. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Throughput does not compensate for poor measurement quality.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Catalyst surfaces can reconstruct, oxidise, foul or form intermediates during reaction. Operando methods observe structure or chemistry under working conditions, helping connect modelled sites to active states. Synchronised reaction data and characterisation reduce the risk of explaining performance using the wrong surface.

Efficient AI matters scientifically because compute, memory and time determine how widely a model can explore and how readily other researchers can reproduce the search. The chapter develops operando evidence literacy. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Build a time-aligned record of condition, structure and product signals. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: The prepared surface may differ from the working active state.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

One target is catalyst discovery for converting carbon dioxide into fuels or chemicals. A useful evaluation includes source and purity of CO₂, energy, conversion, product selectivity, catalyst lifetime and the fate of carbon. A pathway can recycle carbon without being low-emission overall.

Catalysis happens at surfaces where atoms adsorb, react and leave; tiny changes in local structure can alter activity, selectivity, stability and the energy needed for a chemical pathway. The chapter develops carbon-accounting literacy. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Create a carbon-and-energy boundary for a hypothetical conversion route. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: CO₂ utilisation is not automatically climate benefit.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

NRF also names converting glycerol, a biodiesel by-product, into higher-value chemicals. Feedstock composition, purification and competing uses affect process value. Catalyst performance should be tested with realistic impurities and product-separation demands, not only clean laboratory mixtures.

The terascale challenge comes from the huge number of possible compositions, structures and sites, making efficient representations and carefully chosen experiments central to discovery. The chapter develops process-systems reasoning. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Map feedstock, reaction, separation, product and waste streams. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A valuable product in principle may not make a viable process.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

The NRF release states an aim to reduce catalyst-identification time and cost by at least five times, while Imperial describes a 500-fold acceleration for a computational screening phase. These refer to different stages and baselines. Reports should name the task, hardware, starting method and downstream validation before comparing figures.

Inverse design begins with desired performance and searches backward for candidates, but synthesis, characterisation and realistic reaction tests must close the loop. The chapter develops claim auditing. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Rewrite an acceleration claim with task, baseline, metric and boundary. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Screening acceleration does not equal full development acceleration.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Teams need chemistry, physics, materials science, mathematics, machine learning, high-performance computing, robotics and process engineering. Students can model a small adsorption dataset, compare a baseline and discuss uncertainty while also learning laboratory safety and measurement. Official providers control current courses and roles.

Efficient AI matters scientifically because compute, memory and time determine how widely a model can explore and how readily other researchers can reproduce the search. The chapter develops digital-chemistry career preparation. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Plan a small catalyst-data project with a chemical baseline and error analysis. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A computational exercise does not demonstrate a working catalyst.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Archive the programme edition, dataset provenance, structure and reaction identifiers, calculation and experimental methods, model and compute versions, splits, candidate-selection logic, synthesis, operando evidence, performance, lifecycle boundary and released conclusion. This makes acceleration and validity independently assessable.

Catalysis happens at surfaces where atoms adsorb, react and leave; tiny changes in local structure can alter activity, selectivity, stability and the energy needed for a chemical pathway. The chapter develops evidence governance. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF, A*STAR and Imperial describe the programme; project teams, laboratories, data owners and industrial partners control current methods, access and validated results.

Put this into practice by turning the idea into a small, reviewable action: Create a claim-to-data-to-experiment lineage with versions and owners. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: The latest authorised project results supersede announced targets.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter