WellScreen: Digital Wellbeing & Well Screening

• WellScreen is a dual-use concept that includes a lightweight digital wellbeing probe and a geoscience method for screening well performance.
• The digital wellbeing probe measures discrepancies between self-estimated and actual smartphone use, improving self-awareness and positive affect.
• In geoscience, WellScreen employs performance indices, geostatistics, and learned surrogates to rank well locations and optimize drilling decisions.

Searching arXiv for the cited WellScreen-related papers to ground the article. WellScreen denotes two closely related but technically distinct ideas in current research. In human–computer interaction, it is the name of a lightweight technology probe that scaffolds daily reflection on smartphone use by juxtaposing estimated and actual behavior, with the explicit aim of supporting digital self-awareness and wellbeing (Bhat et al., 26 Sep 2025). In subsurface engineering and geoscience, the same label is used more broadly for workflows that screen, rank, monitor, or locate wells by combining performance indices, geostatistics, learned surrogates, physics-informed models, remote sensing, and electromagnetic forward modeling (Lu, 2022). This dual usage situates WellScreen at the intersection of decision support, measurement, and model-based screening.

1. Terminological scope and conceptual structure

The explicitly named WellScreen system is a web-based assistant for digital wellbeing. It is described as a technology probe whose central mechanism is to make users regularly estimate their smartphone use, confront the actual use recorded by the device, and reflect on discrepancies between the two (Bhat et al., 26 Sep 2025). Its conceptual lever is the estimated–actual gap, denoted by a relative difference such as

$$\Delta\%_{c,d} = \frac{E_{c,d} - A_{c,d}}{A_{c,d}} \times 100,$$

where negative values indicate underestimation and positive values indicate overestimation.

In subsurface research, “WellScreen” functions less as a single named product and more as a design motif for screening and ranking well-related decisions. The most explicit geoscience formulation appears in the use of a Well Performance Index (WPI) for selecting additional well locations in the Cana Woodford shale, where kriging and co-kriging are used to generate field-wide maps of predicted producing potential (Lu, 2022). Related work extends this screening logic to portfolio-wide operational intelligence, graph-network surrogates for well placement and control, remote-sensing discovery of wells from satellite imagery, and high-fidelity forward models for steel-cased-well monitoring (Rebello et al., 26 Apr 2026, Tang et al., 2023, Seth et al., 2024, Heagy et al., 2018, Walter et al., 12 Jul 2025).

This suggests a broader interpretation: WellScreen is not a single method but a family of screening systems in which heterogeneous measurements are transformed into ranked, interpretable, and operationally useful representations.

2. WellScreen as a digital wellbeing probe

WellScreen, in the HCI sense, is a lightweight, web-based assistant with six core screens: Login screen, Start-of-day (SoD) estimation, End-of-day (EoD) estimation, Actual report, Visualization, and Reflection survey (Bhat et al., 26 Sep 2025). The daily interaction is organized around five app categories—Creativity, Entertainment, Productivity, Shopping, and Social—for which participants enter SoD estimates, EoD recollections, and actual usage from the phone’s built-in screen time dashboard. The system then displays a comparative bar chart showing SoD prediction, EoD recollection, and Actual usage side by side.

The probe was evaluated in a 14-day in-the-wild deployment with 25 participants; 20 completed the full protocol (Bhat et al., 26 Sep 2025). Participants often underestimated productivity and social media while overestimating entertainment app use, and the paper reports a 10% improvement in positive affect. In the detailed results, median positive affect increased from 29 at entry to 34 at exit, the average change was +10.03%, the paired t-test was $t = 2.63, p<0.05$, and the effect size was Cohen’s $d = 0.63$. Negative affect changed by about −0.12% and was non-significant.

The quantitative misestimation patterns were category-specific rather than purely aggregate. For Entertainment, the actual mean was 104.64 minutes/day, the SoD estimate was 118 min and the EoD estimate 112.27 min. For Productivity, the actual mean was 69.04 min/day, the SoD estimate 59.13 min, and the EoD estimate 62.13 min. For Social, the actual mean was 165.02 min/day, the SoD estimate 150.72 min, and the EoD estimate 167.00 min. The study also reports that self-control (BSCS) was consistently associated with smaller estimated–actual gaps, with coefficients such as $\beta = -16.93, p<0.01$ for the aggregated SoD gap and $\beta = -11.04, p<0.05$ for the aggregated EoD gap (Bhat et al., 26 Sep 2025).

Usability and intervention fit were moderate rather than maximal. The System Usability Scale (SUS) had median = 80 and mean ≈ 71.9, while the Intervention Appropriateness Measure (IAM) had median = 14 and mean ≈ 13.24 on a 4–20 scale (Bhat et al., 26 Sep 2025). Interviews indicated that structured reflection supported recognition of patterns, adjustment of expectations, and more intentional engagement with technology. The system therefore frames discrepancy not primarily as error to be eliminated, but as a prompt for self-explanation and self-awareness.

3. Geostatistical WellScreen for well-location screening

In the geostatistical setting, WellScreen is anchored by the Well Performance Index (WPI), introduced as a single quantitative indicator of how much producing potential a well has, based on early production and pressure behavior (Lu, 2022). In its original form,

$$\text{WPI} = \sum_{i=1}^{90} \text{dailyProdRate}_i \times \text{dailyPressure}_i.$$

The simplified estimator used in the Cana study assumes constant pressure over the first 90 days and approximates fracture pressure by fracture gradient × true vertical depth (TVD):

$$\widehat{\text{WPI}} = \bar{q}_{90} \cdot 90 \cdot G_{\text{frac}} \cdot \text{TVD}.$$

Here $\bar{q}_{90}$ is the average production rate over the first 90 days.

The study uses Cana Field (Cana Woodford Shale, Oklahoma), with over 400 shale gas wells in the completion database and 190 wells with WPI samples used in the geostatistical modeling (Lu, 2022). The primary regionalized variable is $\log_{10}(\text{WPI})$. Secondary variables for co-kriging are $\log_{10}(\text{Fluid volume})$ and $t = 2.63, p<0.05$0, with reported correlations of WPI vs Fluid volume: Pearson correlation ≈ 0.58 and WPI vs Proppant volume: correlation ≈ 0.37.

The interpolation backbone is ordinary kriging and co-kriging. Ordinary kriging predicts

$t = 2.63, p<0.05$1

while co-kriging augments the estimate with a correlated secondary variable,

$t = 2.63, p<0.05$2

The paper fits direct variograms for $t = 2.63, p<0.05$3, $t = 2.63, p<0.05$4, and $t = 2.63, p<0.05$5, as well as cross-variograms between WPI and each secondary variable, enforcing a linear model of coregionalization in which direct and cross-variograms share the same shape and range, but have different partial sills and nuggets.

Model comparison uses leave-one-out (LOO) cross-validation over all 190 WPI samples. The reported metrics are Mean Error (ME) and Root Mean Square Error (RMSE):

$t = 2.63, p<0.05$6

The numerical results are:

Model	ME	RMSE
OK	−0.0014	0.249
CK with fluid volume	−0.0007	0.232
CK with proppant volume	−0.0009	0.229

The paper states that co-kriging with clean fluid volume has the best performance, while the detailed comparison notes that CK with proppant has slightly lower RMSE than CK with fluid, but that the fluid-based CK is preferred because its variogram modeling better honors the experimental variogram (Lu, 2022). The resulting maps identify “sweet spots,” recommend drilling in the “dollar sign” area, and mark lower-potential “warning” zones. In this usage, WellScreen is fundamentally a spatial ranking system: high predicted WPI indicates preferred drilling targets, especially when coupled with low kriging variance.

4. Learned surrogates and design-aware WellScreen systems

Several recent models generalize the screening idea from static maps to dynamic, portfolio-scale prediction. The most explicit is WISE-FM (Well Intelligence and Systems Engineering Foundation Model), a design-aware, physics-informed multi-task model that conditions operational embeddings on static well design by combining Feature-wise Linear Modulation (FiLM), cross-modal attention, multi-task learning, and structural mass conservation with soft physics constraints (Rebello et al., 26 Apr 2026). On the ManyWells benchmark (2000 simulated wells, $t = 2.63, p<0.05$7 data points), the paper reports that design-aware models reduce VFM prediction error by up to $t = 2.63, p<0.05$8 compared to design-unaware baselines, that physics constraints reduce negative flow predictions by 65%, and that flow regime classification achieves 97.7% bottomhole accuracy. On five Equinor Volve producers, the transferred model achieves oil rate $t = 2.63, p<0.05$9, bottomhole pressure $d = 0.63$0, and water rate $d = 0.63$1, and serves as a fast surrogate for integrity-aware well design optimisation over a 24-dimensional design space, with more than $d = 0.63$2 speedup over drift-flux simulations.

The WISE-FM mapping is explicitly multi-task:

$d = 0.63$3

where $d = 0.63$4 is a static design vector and $d = 0.63$5 is a time-varying operational vector. Structural mass conservation is enforced by computing total flow as

$d = 0.63$6

rather than predicting it independently. This makes WellScreen, in this form, a portfolio-wide digital twin linking design, operation, and integrity-related flow regime classification.

A second line of work uses graph surrogates for optimization. The Graph Network Surrogate Model (GNSM) transforms the flow model into a computational graph with an encoding-processing-decoding architecture, constructs separate networks for pressure and saturation, and enhances performance through the inclusion of the single-phase steady-state pressure solution as a feature (Tang et al., 2023). In a 2D unstructured model of a channelized reservoir, with five injection wells and five production wells placed randomly throughout the model and random bottom-hole pressure controls, the model attains median relative error in pressure and saturation for 300 such test cases is 1–2%, produces optimization results comparable to simulation-based optimization, and delivers a runtime speedup of a factor of 36. The paper reports ~120 seconds for a single ADGPRS run versus ~3.3 s per GNSM evaluation.

A third strand focuses on near-well transients. WellPINN addresses the failure of standard PINNs near sharp well singularities by decomposing the domain into stepwise shrinking subdomains with a simultaneously reducing equivalent well radius and combining multiple sequentially trained PINNs (Walter et al., 12 Jul 2025). In the reported case, three sequential PINNs with $d = 0.63$7 reduce the maximum AE from 0.53 in the first-stage solution to 0.11 in the third-stage solution, reduce the maximum AR from 11 to 0.31, and achieve MAE = $d = 0.63$8 while resolving a 10 cm well in a 100 m domain. This makes WellScreen plausible not only as a ranking layer, but also as a near-well forward or inverse engine for operational scenario simulation.

Taken together, these models indicate that a learned WellScreen stack can operate at three levels: field-scale ranking, portfolio-scale operational monitoring, and near-well transient inference.

5. Remote-sensing WellScreen and registry-scale well discovery

A different research direction treats WellScreen as a well-location system based on satellite imagery. The Alberta Wells Dataset introduces a large-scale benchmark for pinpointing oil and gas wells from PlanetScope imagery, with over 213,000 wells (abandoned, suspended, and active) verified against the Alberta Energy Regulator registry (Seth et al., 2024). The final dataset contains 213,447 wells in 188,688 image patches, split into 167,436 train, 9,463 validation, and 11,789 test patches. The well-state counts are Suspended: 55,007, Abandoned: 54,947, and Active: 107,139.

The imagery is PlanetScope PSB.SD, with ground sample distance (nadir): 3.7–4.2 m/px, four spectral bands—Blue (465–515 nm), Green (547–585 nm), Red (650–680 nm), and NIR (845–885 nm)—and patches of 1050 m × 1050 m (Seth et al., 2024). Labels include binary segmentation masks, multi-class segmentation masks for Active / Suspended / Abandoned, COCO-format bounding boxes, and patch-level metadata such as wells_present and no_of_wells. The dataset uses a geographically stratified split based on a two-level K-Means clustering procedure designed to reduce spatial leakage.

Baseline segmentation and detection results establish the difficulty of registry-scale WellScreen. For binary segmentation, U-Net + EfficientNetB6 attains IoU: $d = 0.63$9, F1: $\beta = -16.93, p<0.01$0, Precision: $\beta = -16.93, p<0.01$1, and Recall: $\beta = -16.93, p<0.01$2, while UperNet (Swin Small) attains Recall: $\beta = -16.93, p<0.01$3 with IoU: $\beta = -16.93, p<0.01$4 (Seth et al., 2024). For object detection, SSD Lite yields IoU$\beta = -16.93, p<0.01$5: $\beta = -16.93, p<0.01$6, and FCOS yields mAP$\beta = -16.93, p<0.01$7: $\beta = -16.93, p<0.01$8. The paper emphasizes significant room for improvement, noting challenges from small object size, vegetation, background heterogeneity, temporal mismatch, and label noise due to missing or misclassified wells.

This work places WellScreen in a regulatory and environmental frame. A plausible implication is that remote-sensing WellScreen can function as a large-area front end: detect candidate wells, compare them against registries, and prioritize undocumented or misclassified sites for field verification and plugging.

6. Physics-based monitoring, steel-cased wells, and forward-modeling backbones

WellScreen workflows that rely on geophysics require accurate forward models around wells, especially steel-cased wells. A finite-volume framework on cylindrically symmetric and 3D cylindrical meshes provides modeling capabilities for direct current resistivity, time domain electromagnetics, and frequency domain electromagnetics with explicit support for variable electrical conductivity and magnetic permeability (Heagy et al., 2018). The governing equations are expressed in quasi-static form, for example in the frequency domain:

$\beta = -16.93, p<0.01$9

with constitutive relations

$\beta = -11.04, p<0.05$0

The framework is motivated by the fact that steel casings have both high conductivity and significant magnetic permeability, creating large property contrasts and a severe disparity in length scales. The implementation is part of the SimPEG software ecosystem, and the paper reports substantial computational savings from cylindrical discretization. In a TDEM validation case, SimPEG 3D cylindrical: 314k cells, 14 min on single core, compared with UBC OcTree FV: 5.0M cells, 57 min on single core, Commer FD: 2.18M cells, 23.2 hours on 512 cores, and Commer FE: 8.4M tetrahedra, 63 hours on single core (Heagy et al., 2018).

The physical analyses are directly relevant to screening and monitoring. In DC resistivity, the paper revisits the near zone, intermediate zone, and far zone structure around a casing, showing how current channelling and charge build-up depend on casing conductance and length. In TDEM, a top-casing extended electrode configuration demonstrates how image currents form after shut-off and how steel casings can increase current density near the casing at depth. In FDEM, comparisons between copper and iron pipes show that responses are not controlled simply by $\beta = -11.04, p<0.05$1; geometry and independent conductivity and permeability effects matter.

Within a WellScreen interpretation, this framework supplies the forward-modeling backbone for survey design, well-logging correction, and monitoring suitability assessment. It can determine whether a well behaves like a useful extended electrode, whether casing effects will mask or enhance a target, and how source–receiver placement should be chosen.

7. Limitations, assumptions, and emerging directions

Across the literature, WellScreen systems are constrained by strong modeling assumptions. In the geostatistical workflow, the WPI estimator assumes constant pressure over the first 90 days and uses frac gradient × TVD as a proxy for pressure; the spatial model assumes second-order stationarity, approximate isotropy, uses $\beta = -11.04, p<0.05$2 transforms, and, for co-kriging, restricts modeling to co-located data only to satisfy the linear model of coregionalization (Lu, 2022). These assumptions make the workflow tractable but limit its direct transfer to settings with strong non-stationarity, explicit anisotropy, or incomplete multivariate sampling.

The learned models inherit distributional limits. WISE-FM is explicitly motivated by the fact that deploying models across diverse portfolios requires generalisation to wells with design parameters outside the training distribution, and its results argue that design awareness, physics enforcement, and multi-task learning are essential and complementary ingredients (Rebello et al., 26 Apr 2026). GNSM, as reported, is currently demonstrated on a 2D incompressible two-phase setting with constant-in-time BHP controls and a training focus on one base geomodel, while WellPINN is demonstrated for a single well in a homogeneous domain and remains sensitive to the choice of $\beta = -11.04, p<0.05$3, collocation sampling, and sequential decomposition depth (Tang et al., 2023, Walter et al., 12 Jul 2025).

The remote-sensing pipeline is limited by registry incompleteness and regional specificity. The Alberta dataset depends on AER records, acknowledges that some wells are missing or misclassified, and is tied to one geography, one imagery source, and a summer-season optical snapshot (Seth et al., 2024). The digital wellbeing probe is limited by a two-week deployment, a predominantly U.S. college-student sample, manual transcription of actual use, and observer effects that may alter behavior (Bhat et al., 26 Sep 2025).

Even with these constraints, a coherent pattern is visible. This suggests that WellScreen is converging toward an architecture in which screening is not a single prediction problem but a composition of measurement, discrepancy detection, uncertainty control, and action support. In digital wellbeing, the discrepancy is between estimated and actual use. In geostatistics, it is between sparse observations and field-wide interpolation. In learned subsurface models, it is between expensive simulators and fast surrogates. In remote sensing, it is between registry records and image-derived detections. In geophysics, it is between idealized well models and casing-aware field responses. The unifying technical theme is the same: construct a compact representation that is sufficiently faithful to support ranking, intervention, or monitoring under operational constraints.