From Data Layers to Decision Questions: A Question-Oriented Interface Model for an Air-Quality Digital Twin in Thailand

Northern Thailand experiences a severe particulate-matter episode every year between December and April, driven by agricultural and forest burning and by smoke crossing national borders [1]. Provincial officers must decide, often within hours, whether to issue a public warning, whom to coordinate with, and which control measure to press for. International guideline values [2] and Thailand’s 24-hour ambient standard of 37.5 µg/m³, against which the interface studied here reports exceedance, supply the thresholds those decisions are judged by. Platforms that support such decisions now routinely embed machine learning: they fuse ground stations, satellite fire detections and numerical weather output, and return concentration forecasts, source attribution and scenario simulations. What determines whether that output actually changes a decision, however, is the interface. Most environmental dashboards are organized by data layer, with one view for stations, another for satellite products and another for meteorology. That structure mirrors how the data arrives rather than how a decision is made, and it leaves the integration work to the user. Established dashboard design guidance and the recent design-space taxonomy of dashboard genres concentrate on layout, metric selection and visual encoding rather than on whether the navigation matches a decision process [3], [4]. A second difficulty is specific to machine forecasts. A forecast is a distribution, but dashboards overwhelmingly display a single number, and a technically correct graphic can still be read as far more certain than it is. The classic demonstration is the hurricane forecast cone, whose center line led the public to overestimate the precision of the track [5], [6]; more generally, authors omit uncertainty because it is effortful and because audiences are assumed to reject it [7]. Guidance for human-AI interfaces explicitly asks designers to make clear how well a system performs [8], and industry pattern collections restate the point [9], [10], but concrete, reported instantiations in a government setting remain uncommon. The northern Thailand air-quality problem itself is well documented [1]. Interface-level studies of the tools that Thai agencies actually use appear to be absent from the literature, and that absence motivates this case study. We make three contributions. First, we describe a question-oriented information architecture in which each screen is named after the decision question it answers, and we relate its ordering to an established model of situation awareness [11]. Second, we report eight interface patterns for communicating machine forecasts to non-expert officials, derived by coding every panel in the implemented interface, together with how widely each pattern is used. Third, we report a structured usability inspection of the implementation and the redesign priorities it produced. We deliberately treat the underlying forecasting model as given; this is a paper about the interface.

A. Dashboards, situation awareness and decision support Dashboard design has a mature practitioner literature focused on visual clarity and the avoidance of decorative clutter [3], and a research framing that treats analyst-facing displays as instruments for reasoning rather than reporting [12]. Sarikaya et al. show that what practitioners call a dashboard spans several distinct genres with different purposes [4]. Endsley’s model of situation awareness distinguishes perceiving elements in the environment, comprehending what they mean, and projecting what will happen next [11]; we use those three levels as the organizing spine of our architecture, because they describe a progression a screen sequence can follow. B. Communicating forecast uncertainty Spiegelhalter et al. survey how predictive uncertainty can be presented to lay audiences and stress that format changes conclusions [5]. Broad et al. document systematic misreading of an official forecast graphic [6], and Hullman explains why practitioners avoid showing uncertainty at all [7]. Encouragingly, Kay et al. show that ordinary users can read discrete uncertainty encodings in an everyday predictive system [13], and Padilla et al. give a cognitive account of how such displays are actually processed under time pressure, which is the condition our users work in [14]. C. Human-AI interaction and calibrated trust Amershi et al. consolidate design guidance for AI-infused systems, including making capability and performance legible to the user [8]; practitioner collections operationalize the same ideas [9], [10]. The goal is not maximal trust but appropriate reliance [15], since automation can be both over-relied upon and abandoned after false alarms [16]. Lim et al. show empirically that particular explanation types raise intelligibility, which motivates co-locating an explanation with the value it justifies [17]. D. Air-quality information for non-expert audiences Huber et al. report an evaluated air-quality visualization for citizens and note that some participants struggled to interpret the graphs [18]. Wu et al. show that the wording used to describe an air-quality category measurably changes both risk perception and protective intention [19], which makes label text a design variable rather than a cosmetic choice.

THE PLATFORM AND ITS USERS The system studied here is an implemented prototype of a national environmental digital twin for PM2.5. It ingests measurements from 174 government monitoring stations, satellite fire detections, numerical weather forecasts and static terrain and land-cover layers, places them on a common national grid covering 77 provinces, and produces hourly concentration forecasts with lead times to seven days, province-to-province source attribution, threshold-based early warning, and what-if simulation of control measures. A physics-informed model provides the forecasts; its formulation is outside the scope of this paper. The intended users are provincial environmental and disaster-prevention officers, not atmospheric scientists. Three properties of that audience drove the design. They are accountable, so they need to be able to justify a decision afterwards. They are time-poor, so a screen must answer one question rather than present a data inventory. And they work bilingually, so every panel title carries an English term with a Thai explanatory line beneath it. The interface is a single-page web application with seven task screens and a settings screen, rendered on a dark background for continuous display in an operations room.

A QUESTION-ORIENTED INTERFACE MODEL The central design decision was to name each screen after the question it answers instead of the data it contains. Table I lists the seven screens, the question each one owns and what it principally presents. Fig. 1 shows the resulting structure. The sequence moves from perceiving the current state, through comprehending where the pollution came from and where it is heading, to projecting what will happen and what an intervention would change, and finally to the two questions a public decision actually turns on: why the system reached its conclusion, and whether its past record justifies acting on it. Fig. 2 contrasts this with the conventional layer-oriented arrangement. Three consequences of the arrangement are worth drawing out. First, the navigation bar itself becomes a checklist of the decision process, so a user who works left to right is intended to arrive at a defensible procedure. Second, because every screen fuses all of the underlying layers, cross-layer integration becomes the system’s responsibility rather than the user’s; the layer inventory survives only as a provenance panel. Third, the ordering follows the perception-comprehension-projection progression [11], so the interface is designed to model the reasoning sequence it supports. Two of the seven questions are unusual in this genre and are worth naming. “Where is it going” inverts source attribution into a forward view, which is what a province needs in order to warn its neighbors rather than merely explain its own air. “Can I trust it” promotes forecast verification out of an internal engineering console and into the user interface, where an accountable official can inspect it. The seven screens share a layout grammar that is visible even at the scale of Fig. 6. A left rail carries the current state of the selected province together with the model’s own confidence in it; a centre column carries the map and, beneath it, the same now-to-seven-day lead-time strip on every screen; and a right rail carries the screen’s answer to its question, closing with a control that hands off to the next question in the sequence, so that the source screen previews a scenario, the impact screen names who will be affected next, and the forecast screen lists who will exceed the standard. The grammar is what makes the question-orientation usable rather than merely nominal: because the frame does not move, the only thing that changes from screen to screen is the answer, and attention is not spent relocating the controls. Each screen resolves its question into a distinct representational form, and it is the choice of form, more than the choice of variable, that constitutes the design. S1 is an overview: the whole country at one glance, with the regional and source breakdowns arranged around the map as context rather than as destinations, so that the opening act is orientation and not query. S2 and S3 render one underlying transport relation as two directed views, incoming and outgoing, switched by a single toggle rather than separated into two screens; each is drawn twice over, as labeled paths on the map for a question that is spatial, and as a ranked list beside it for a reader who has to name and cite a contributor. S4 moves the representation from space to time: the map persists, but the primary object becomes a trajectory with a marked peak and a band around it, and the exceedance question is answered separately as a discrete watchlist, because crossing a threshold is a categorical event rather than a point on a curve [14]. S5 is the only screen that takes a substantive input; the user composes a package from a checklist and the display answers as a before-and-after comparison that keeps the baseline in view, so that the effect of an intervention is read as a difference rather than reported as a figure. S6 places the account next to the value, in the pattern of why and why-not explanation [17], and supplies it twice over: as continuous prose for an officer who must write a justification, and as a ranked driver list for one who must check a particular factor. S7 turns the verification record into a user-facing artifact rather than an engineering console, pairing the aggregate metrics with a province-by-horizon matrix, so that the question an accountable reader actually has — is it accurate

All seven screens are implemented and running as a single-page web application, and Fig. 6 reproduces them, one panel per screen, in the order in which the navigation presents them. The capture is the material from which the panel coding of Section VI and the inspection of Section VII were made, and it is reproduced here so that the claims we make about the interface can be checked against the interface itself. Fig. 7(a) shows the most consequential of the disclosure devices as it is actually rendered. A banner sits across the map on the forecast, scenario and analysis screens and states that the model returns no dispersion value and that the band shown is therefore an approximation. The verification screen is likewise not an internal console but a user-facing report centre with seven templates and export in five formats, and every derived figure on it carries an estimate tag. Volunteering a limitation on the screen where it bears is what the guideline literature asks for [8] and what practitioners are documented as avoiding [7]. here, and at the lead time on which I intend to act — can be put to the display directly. Two properties of the capture should be stated plainly. It is one hour, so it fixes the interface but not its behavior across the range of conditions the system must handle; and it falls in August, outside the December-to-April episode the system is built for, so concentrations are low and no alarm state is exercised. Neither bears on the panel coding or on the inspection, both of which concern how a quantity is presented rather than which quantity happens to be present, but Fig. 6 should not be read as a picture of the system under load.

To characterize the interface systematically rather than anecdotally, we enumerated the panel regions of the seven screens of Fig. 6 and coded each for the design devices it uses. We counted one region per named panel identifiable in the captured viewport of a screen, including regions only partly in view, and counted a control shared between screens once for each screen on which it appears. That gives 97 panels, between 12 and 16 per screen. Eight recurring devices emerged, listed in Table II: 91 of the 97 panels instantiate at least one, and six appear on all seven screens. Fig. 3 gives the layout of one screen with the patterns marked, Fig. 4 a detail of that screen as implemented, and Fig. 5 the coverage across the interface. The most pervasive pattern is also the cheapest. Every technical quantity is accompanied by a short sentence stating what it means for a decision: a root-mean-square error is glossed as weighting large misses more heavily, and an outbound-influence score is glossed as the chance that this province sends dust to its neighbors. Given that category wording measurably shifts risk perception [19], these glosses are substantive rather than decorative. Confidence disclosure and provenance tagging between them cover 37 distinct panels of the 97 and are the interface’s main answer to the over-confidence problem. Values are marked as measured, modeled, interpolated or estimated; lead times that were interpolated between published forecast steps are flagged in the control strip itself. Most strikingly, the forecast, scenario and analysis screens carry a banner stating that the model does not emit a predictive spread and that the displayed band is therefore approximate. Volunteering a limitation of exactly this kind is what the guideline literature asks for [8] and what practitioners are documented as avoiding [7]. Two further patterns deserve comment. Scalar decision proxies reduce a province-to-province transport matrix to two bounded scores, one for influence exerted and one for influence received, each with a plain definition; this is a deliberate loss of information in exchange for a quantity a non-specialist can act on. Threshold anchoring ties color and band boundaries to the national 24-hour standard rather than to data quantiles, so that a change of color always means a change of regulatory status. The band edges are not identical across screens, however, which the evaluation records as a finding.

A. Method We inspected all seven screens against Nielsen’s usability heuristics using the heuristic-evaluation procedure [20], [21], and additionally against the human-AI interaction guidelines concerning legibility of capability and performance [8]. Each finding received a severity rating on Nielsen’s scale [22]. We then ran a task-based walkthrough [23] of three representative tasks: T1, decide whether to issue a public warning today; T2, establish whether the smoke affecting a province originated outside it, in order to request cross-border coordination; and T3, justify to a superior that recent forecasts were reliable enough to have acted on. The inspection was carried out by the design team rather than by independent evaluators, so the severities should be read as a priority ordering rather than as calibrated estimates. B. Results The walkthrough indicated that T2 and T3 are well served: each maps onto a single screen and completes without leaving it, which is the intended payoff of the architecture. T1 does not. Deciding whether to warn requires the current state from S1, the exceedance outlook from S4 and the ranked action from S6, so the most consequential task is the one the architecture serves least directly. This produced finding F11. Fourteen findings emerged in total, listed in Table III: six major, five minor and three cosmetic. They cluster into two systemic issues rather than fourteen unrelated defects. The first is that the interface presents its numbers more precisely than its own evidence supports. Concentrations appear to 0.1 µg/m³ while the platform’s own verification screen reports a mean absolute error of 2.1 µg/m³, a mixing-layer height of zero is displayed as though it were a measurement rather than an absent value, and two map legends are labeled in milligrams where the table beside them uses micrograms. All three invite exactly the over-reliance the uncertainty patterns were introduced to prevent [15]. Fig. 7(b)–(d) show the three as they are rendered. The second is that the interface’s candor is inconsistently executed. It discloses component scores without their scale, so a weighted share of 40 percent sits beside a data-quality score of 86 percent in identical visual form and reads as a poor grade; a confidence label carries three different values on different screens; band edges differ between screens; and monetized health benefits appear in the same tile style as measured concentrations. Disclosure that a reader cannot interpret can depress trust as readily as absent disclosure, and abandonment is as real a failure mode as over-reliance [16]. Fig. 7(e)–(g) show three of them. Two further major findings concern a quantity that takes different values on different screens, and they emerged from reading the capture along a decision path rather than screen by screen. Fig. 8(a) and (b) give the source decomposition for one province at one hour as the source screen and the analysis screen each present it: the first attributes 85 percent of the concentration across six named origins and leaves the remainder unaccounted for, the second renormalizes the same six to 99 percent, and the leading contributor accordingly reads 38 percent on one screen and 44 percent on the other, with neither screen stating its basis. Fig. 8(c) is sharper. A header tile reports a population at risk of zero, qualified as not yet above threshold, while a callout on the same map puts roughly 195,000 people at risk within 42.9 hours; both are labeled as population at risk, and an official who reads the header and stops there would stop for the wrong reason. Defects of this kind are invisible to a heuristic pass over one screen at a time, which we record as an argument for inspecting a decision-oriented interface along its decision path as well as against its heuristics.

Four implications may generalize beyond this system. Name screens after questions, because a navigation bar that reads as a decision procedure is itself a form of guidance. Gloss every technical quantity, because the gloss costs one line and determines whether the quantity is used at all [19]. Distinguish measured from derived values everywhere, because a dashboard that mixes them silently transfers the model’s uncertainty onto the official’s accountability. And anchor encodings to regulatory thresholds rather than to the data, so that visual change carries institutional meaning. A fifth implication is methodological. Because a question-oriented architecture distributes one decision across several screens, it has to be inspected along the decision path as well as screen by screen: the defects that survive a per-screen heuristic pass are precisely those in which two screens disagree. The architecture is not specific to air quality. Flood, heat-wave and epidemic decision support pose the same seven questions, and only the domain-specific contents of each screen change. What does not transfer automatically is the balance between candor and usability. Our evaluation found the disclosure patterns implemented but inconsistently executed, and disclosure that a user cannot interpret risks producing disuse rather than calibrated reliance [16]. Uncertainty patterns therefore need the plain-language pattern applied to them as well, not only to the point estimates. We also record the bilingual labeling scheme as a design decision rather than a validated one. Every panel title pairs an English term with a Thai explanatory line, which keeps internationally standard technical terms intact while carrying the interpretation in the users’ working language. Culturally adapted interfaces can improve performance and perceived usability [24], but we have not tested this scheme and claim no conformance to localization guidance.

The most important limitation is that no end-user study has yet been conducted. The evaluation reported here is an expert inspection by the design team, with the consequences for severity calibration noted above, and it cannot establish that officials reach better or faster decisions. A field study with provincial officers is planned for the next burning season, measuring task completion, decision confidence and whether the disclosures are read as intended. Second, the coding was performed by the authors on a single system, so the eight patterns are a candidate vocabulary awaiting independent application. Third, the case is one domain in one country, the captured state is a single off-season hour, and the walkthrough covered only three tasks; a fuller inventory would probably surface further gaps of the kind F11 represents.

Organizing an environmental decision-support interface around seven decision questions, rather than around the data layers that feed it, moves the work of integration from the user to the system and turns the navigation into a record of a defensible procedure. Coding all 97 panels of an implemented PM2.5 digital twin yielded eight interface patterns for presenting machine forecasts to accountable non-experts, of which decision-question framing, plain-language glossing and confidence disclosure are the most widely instantiated. A structured inspection indicated that the harder problem is not adopting such patterns but executing them consistently: all six major findings concern an uncertain quantity presented with unwarranted precision, in the wrong unit, without the qualification a reader needs, or with a different value on a different screen. Interfaces that ask officials to act on machine forecasts must make the limits of those forecasts as easy to read as the forecasts themselves.

This work is part of a research project supported by the National Research Council of Thailand (NRCT) under grant no. N25B(PM)690041. The authors gratefully acknowledge this support, and thank the agencies and provincial officers whose operational practice informed the design of the interface studied here. The views expressed are those of the authors and do not necessarily reflect those of the funding agency.