These definitions are for annotators, reviewers and extraction prompts. Each type has a boundary test, because the hard part of labeling a paper is deciding between neighboring types. The classes are in ontology/schema/seismohydro.linkml.yaml, and seed instances are in ontology/vocab/.
All of this is a proposal for review in M3. Types that annotators can't apply consistently, or that abstracts rarely mention, should be dropped or curated by hand.
| Type | Definition | Boundary test | Examples |
|---|---|---|---|
| Sensor | A physical instrument or platform that records a signal | Can you deploy, install or launch it? | Broadband seismometer, DAS, piezometer, stream gauge, GNSS, GRACE |
| Observable | A quantity measured or derived from a record | Is it reported from a record, even if it stands in for something else? Link it to what it stands for with proxy_for |
dv/v, seismic noise power, streamflow, well water level |
| Dataset | A named collection of records, or the network or archive that provides them | Could you download it or request it by name? | FDSN archives, USGS NWIS |
| Type | Definition | Boundary test | Examples |
|---|---|---|---|
| StateVariable | A physical quantity describing the system, measurable directly or not. Includes slowly varying material properties | Would it have a value even with no instrument present? | Pore pressure, saturation, soil moisture, hydraulic head, permeability, strain |
| Compartment | A part of the Earth system where water resides or moves and elastic waves can act | Can you point to where it is in a cross section? | Surface water (river channel, lake or reservoir, wetland), groundwater (saturated zone, vadose zone), snowpack, frozen ground, fault zone |
| Process | A physical mechanism or change through time, described without regard to harm | Does it name something happening, with no reference to damage? | Recharge, infiltration, pore pressure diffusion, bedload transport, earthquake rupture |
| Hazard | A process considered for its potential harm to people, property or ecosystems | Is the study concerned with harm or risk? If not, use Process | Flood, drought, landslide, liquefaction, ground shaking |
Soil moisture is a StateVariable located in the vadose zone, not a compartment. The word "groundwater" can mean the system or the water in it: tag the compartment shared:groundwater, and use a state variable such as hydro:groundwater_storage or hydro:hydraulic_head when a quantity is meant.
| Type | Definition | Boundary test | Examples |
|---|---|---|---|
| Task | The inferential goal of a study | What is the study trying to find out, regardless of how? | Detection, monitoring, forecasting, parameter estimation, imaging, attribution, downscaling |
| Method | A procedure applied to data or to a model to produce a result | Could you write its steps as an algorithm? | Ambient-noise cross-correlation, baseflow separation, data assimilation, LSTM rainfall-runoff |
| PhysicalModel | A mathematical statement of physics, independent of code or data | Can you write it as equations? | Elastic wave equation, Richards equation, Darcy's law, Gassmann fluid substitution |
| Software | A named code that implements a method or model | Does it have a repository or a release? | MODFLOW, ParFlow, SPECFEM, ObsPy |
| ResearchApproach | Whether data, governing equations or both produce the main result | See below | Data-driven, model-driven, hybrid |
A neural network is a Method, not a PhysicalModel, however well it fits.
ResearchApproach is judged on the main result of a study, not on its tools or the words its authors use. "Data-driven" often just means machine learning in abstracts, and it means different things across the two fields. Ask two questions:
- Are governing equations solved or imposed in producing the main result?
- Do records constrain the estimate that is the main result, beyond setting up or checking a model?
| Equations? | Records constrain the result? | Value | Examples |
|---|---|---|---|
| No | Yes | Data-driven | Regressing dv/v on rainfall, LSTM streamflow prediction, deep-learning phase picking, baseflow separation |
| Yes | No | Model-driven | Forward wave simulation, a poroelastic analytical solution, MODFLOW scenario runs with fixed parameters |
| Yes | Yes | Hybrid | Full waveform inversion, ensemble Kalman filtering, physics-informed networks, fitting a poroelastic model to dv/v |
Cases annotators should expect to argue about:
- A calibrated process model used for scenarios is model-driven: the equations produce the result, and calibration only sets it up. If the calibrated parameters are themselves the result, the study is hybrid.
- A measurement study that interprets its result with physics afterward is data-driven, unless a physical model is actually fitted.
- Surrogates trained only on simulations are model-driven, since no records constrain them. If they are also trained on observations, they are hybrid.
- Reviews and method papers with no single main result get no approach value.
Each work gets one primary value through a uses_approach edge. A Method can carry a typical_approach, but the value on the work takes precedence. For example, data assimilation is typically hybrid.
This axis is separate from the kind of evidence. A data-driven study can run on simulated data, and a model-driven study can be checked against field data. Evidence type (observation, lab, simulation, theory) will be an attribute on claims, and isn't implemented yet.
Why it matters for opportunities. It extends the O2 method transfer class. For the same task and variable, one field may work mostly data-driven while the other remains mostly model-driven, and that imbalance suggests what could transfer. It also shows hybrid methods where one field already has both ingredients and hasn't combined them.
| Type | Definition | Boundary test | Examples |
|---|---|---|---|
| Gap | A kind of limitation or unmet need stated in the literature | Does the paper say something can't yet be done or remains unknown? | Sparse observations, poor depth resolution, non-uniqueness, scale mismatch, no field validation |
| Setting | A kind of environment where a study could be repeated | Could it exist in many places? | Karst terrain, arid basin, agricultural basin, mountain catchment |
| Region | A named geographic place | Could you find it on a map by name? | California Central Valley |
| Event | A specific occurrence with a date and a place | Does it have a date? | 2011 Tohoku-oki earthquake |
| DecisionDomain | A sector of human decisions that research can serve | Is it a broad area of practice? | Water resource management, disaster management, subsurface energy and storage operations |
| Decision | A specific choice someone must make, with stateable information needs | Could you name who decides and how soon they need to know? | Groundwater allocation, flood warning, dam safety assessment, fluid injection regulation |
"Application" is deliberately not a type, because it is too vague to annotate consistently. Split it into the Task (monitoring) and the Decision it serves (groundwater allocation).
Gap instances here are categories. A useful gap also needs a subject and a place, such as sparse observations of soil moisture in an arid basin. That requires a reified record joining several nodes, which M3 still has to design.
Decision requirements such as latency and accuracy stay empty until they have a cited source. Abstracts will probably rarely state them, so Decision may end up as a short curated list and not an extracted type.
| Edge | From | To | Status |
|---|---|---|---|
broader |
Any concept | Parent concept | Built, curated |
proxy_for |
Observable | StateVariable | Built, curated seeds |
typical_approach |
Method | ResearchApproach | Built, curated seeds |
uses_approach, uses_method_on, addresses_task, measures, reports_gap, studied_in, informs_decision |
Work or concept | Concept | Proposed, from extraction in M4 |