arxiv: 2604.23767 · v1 · submitted 2026-04-26 · 💻 cs.LG

Recognition: unknown

WISE-FM:Operation-Aware, Engineering-Informed Foundation Model for Multi-Task Well Design

Carine de Menezes Rebello , Anderson Rapello dos Santos , Idelfonso B. R. Nogueira

Authors on Pith no claims yet

Pith reviewed 2026-05-08 06:17 UTC · model grok-4.3

classification 💻 cs.LG

keywords virtual flow meteringphysics-informed machine learningfoundation modelmulti-task learningwell design optimizationflow regime classificationsurrogate modelingoil and gas engineering

0 comments

The pith

A design-aware and physics-informed foundation model reduces virtual flow metering errors by up to 13 times and speeds up well design optimization by over 1000 times.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces WISE-FM, a multi-task model that conditions predictions of flow rates and conditions on well design parameters. It does this through Feature-wise Linear Modulation and cross-modal attention while adding multi-task heads for rates, pressures, and flow regimes plus soft physics constraints enforcing mass conservation. This combination lets one model handle diverse wells instead of building separate models for each. A sympathetic reader would care because ignoring design effects or physical laws currently forces retraining or produces invalid outputs when wells differ from training data. The benchmark results show large accuracy gains on simulated wells and successful transfer to real field data.

Core claim

WISE-FM integrates design conditioning via FiLM and attention, multi-task learning for flow rates, bottomhole conditions, and flow regime classification, and soft physics constraints derived from mass conservation. On the ManyWells benchmark of 2000 simulated wells and one million points, design awareness cuts virtual flow metering error by up to 13 times compared with design-unaware baselines, physics constraints cut negative flow predictions by 65 percent, and flow regime accuracy reaches 97.7 percent. The same model transfers to real data from five Equinor Volve producers with R-squared values of 0.89 for oil rate, 0.98 for bottomhole pressure, and 0.97 for water rate. It also serves as a

What carries the argument

The WISE-FM model that uses Feature-wise Linear Modulation (FiLM) and cross-modal attention to condition operational embeddings on well design parameters, together with multi-task prediction heads and soft physics constraints enforcing mass conservation.

If this is right

Design conditioning enables accurate predictions for wells whose design parameters were never seen during training.
Soft physics constraints reduce physically invalid outputs such as negative flow rates by 65 percent without extra sensors.
Multi-task learning produces simultaneous forecasts of flow rates, bottomhole conditions, and flow regimes from one model.
The trained model replaces drift-flux simulations for optimization over a 24-dimensional design space at more than 1000 times the speed.
High transfer accuracy on real Equinor data shows the approach works on operational measurements beyond simulation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same conditioning on design parameters and soft constraints could apply to other engineered systems whose geometry affects physical behavior, such as pipelines or reactors.
The 1000-fold speedup opens the possibility of embedding the model inside real-time control loops for adaptive well operations.
Extending the model with uncertainty estimates on its multi-task outputs would let operators decide which wells still need physical sensors for verification.

Load-bearing premise

The ManyWells simulated dataset from 2000 wells and one million points sufficiently represents the range and distribution of real-world well designs and operations, and the chosen soft physics constraints are correctly formulated and weighted.

What would settle it

Evaluating the model on a fresh set of real wells whose design parameters fall outside the ManyWells distribution and finding that error reductions disappear or that the physics constraints raise overall error.

Figures

Figures reproduced from arXiv: 2604.23767 by Anderson Rapello dos Santos, Carine de Menezes Rebello, Idelfonso B. R. Nogueira.

**Figure 1.** Figure 1: WISE-FM architecture. Well design parameters enter through two pathways: FiLM conditioning generates scale (γ) and shift (β) vectors that modulate the operational sequence embedding (broadcasting across all T time steps), and cross-modal attention allows each operating point to selectively attend to relevant design features. Total mass flow rate is derived structurally from the three predicted phase compon… view at source ↗

**Figure 2.** Figure 2: Predicted vs. true values for all six prediction targets across 200 held-out test view at source ↗

**Figure 3.** Figure 3: Flow regime confusion matrices for bottomhole (left) and wellhead (right) lo view at source ↗

**Figure 4.** Figure 4: Flow regime transitions along operating curves for four representative test wells. view at source ↗

**Figure 5.** Figure 5: Time-series predictions for all 5 Volve producing wells using the view at source ↗

**Figure 6.** Figure 6: Predicted vs. measured values for all five Volve targets ( view at source ↗

**Figure 7.** Figure 7: Pareto front from the bi-objective design optimization (blue circles with connect view at source ↗

**Figure 8.** Figure 8: Tri-objective Pareto surface: oil production vs. slug-churn probability vs. de view at source ↗

**Figure 9.** Figure 9: Design parameter sensitivity: solid lines show oil production rate (left axis), view at source ↗

**Figure 10.** Figure 10: Integrity risk map: slug-churn probability at bottomhole vs. key design pa view at source ↗

read the original abstract

Deploying machine learning models across diverse well portfolios requires generalisation to wells with design parameters outside the training distribution. Current data-driven approaches to virtual flow metering (VFM) and bottomhole estimation typically treat each well independently or ignore the influence of well design on operational behaviour. We present WISE (Well Intelligence and Systems Engineering Foundation Model), a design-aware, physics-informed multi-task model that integrates three complementary mechanisms: Feature-wise Linear Modulation (FiLM) and cross-modal attention to condition operational embeddings on well design parameters; multi-task learning for simultaneous prediction of flow rates, bottomhole conditions, and flow regime classification; and structural mass conservation with soft physics constraints derived from well engineering principles. Evaluation on the ManyWells benchmark (2000 simulated wells, $10^6$ data points) demonstrates that design-aware models reduce VFM prediction error by up to $13\times$ compared to design-unaware baselines, and that physics constraints reduce negative flow predictions by 65%. Flow regime classification achieves 97.7% bottomhole accuracy, providing continuous well integrity monitoring without additional sensors. The methodology transfers to real operational data from five Equinor Volve producers (oil rate $R^2 = 0.89$, bottomhole pressure $R^2 = 0.98$, water rate $R^2 = 0.97$). The trained model additionally serves as a fast surrogate for integrity-aware well design optimisation over a 24-dimensional design space, with more than $1000\times$ speedup over drift-flux simulations. These results demonstrate that design awareness, physics enforcement, and multi-task learning are essential and complementary ingredients for foundation models intended to operate across large well portfolios.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Design conditioning and physics constraints deliver clear wins on simulated wells, but real-data validation lacks the necessary controls and comparisons.

read the letter

The main thing to know is that this paper gets solid gains from conditioning on well design and adding physics constraints, but only on simulated data. The real data results are too thin to confirm the generalization story. What stands out as new is the combination of FiLM modulation and cross-modal attention to make the model aware of design parameters like tubing size or perforation details, paired with multi-task learning across flow rates, bottomhole conditions, and regime classification, all wrapped with soft mass conservation constraints from engineering principles. On the ManyWells benchmark of 2000 simulated wells, this cuts virtual flow metering error by 13 times compared to unaware models and reduces invalid negative flow predictions by 65%. The model also doubles as a quick surrogate for optimizing designs in 24 dimensions, running over 1000 times faster than full drift-flux simulations. Those numbers are concrete and the setup seems reproducible from the description. The soft spot is the real-world part. They show R-squared values of 0.89 to 0.98 on five Volve wells for oil rate, pressure, and water rate, but skip any comparison to baselines or ablations on that data. Without that, it's hard to say the design awareness or constraints are doing the heavy lifting outside simulation. Real wells bring sensor noise and distribution shifts that the simulator might not capture, so the claimed benefits could shrink. This paper is for people building ML tools for oil and gas operations, particularly those scaling models across many wells. A reader working on similar surrogate modeling or physics-informed networks would get some practical ideas from the architecture choices. I would send it to peer review. The simulated evidence is strong enough to warrant referee input on the methods and whether the real-data gap can be addressed.

Referee Report

4 major / 2 minor

Summary. The paper presents WISE-FM, a design-aware, physics-informed multi-task foundation model for virtual flow metering (VFM) and well design. It integrates FiLM conditioning and cross-modal attention to incorporate well design parameters into operational embeddings, multi-task learning to predict flow rates, bottomhole conditions, and flow regimes simultaneously, and soft physics constraints enforcing mass conservation from well engineering principles. On the ManyWells simulated benchmark (2000 wells, 10^6 points), it claims up to 13× VFM error reduction versus design-unaware baselines and 65% fewer negative flow predictions with physics constraints; flow regime classification reaches 97.7% accuracy. The model transfers to real Equinor Volve data (five producers) with R² values of 0.89 (oil rate), 0.98 (bottomhole pressure), and 0.97 (water rate), and serves as a surrogate enabling >1000× faster integrity-aware optimization over a 24-dimensional design space compared to drift-flux simulations.

Significance. If the quantitative claims and generalization hold, this represents a meaningful advance in applying foundation models to petroleum engineering by addressing design dependence and physical consistency in VFM and well optimization. The large-scale simulated benchmark, multi-task formulation, and attempt at real-data transfer are strengths that could influence surrogate modeling practices. The 1000× speedup potential for design optimization is notable if surrogate fidelity is maintained.

major comments (4)

[§4.2 (ManyWells evaluation)] §4.2 (ManyWells evaluation): The central 13× VFM error reduction and 65% negative-flow reduction claims lack specification of the precise error metric (MAE/RMSE/etc.), the exact design-unaware baseline architectures, data splits or cross-validation procedure, and any error bars or statistical tests across the 2000 wells. These omissions are load-bearing because they prevent independent verification of whether design awareness and physics constraints deliver the reported gains.
[§4.3 (real-data transfer)] §4.3 (real-data transfer): Performance on the five Volve producers is reported solely as R² values with no baseline comparisons, no ablations removing FiLM conditioning or the physics constraints, and no additional metrics (e.g., MAE or negative-flow counts). This is load-bearing for the generalization claim, especially given potential distribution shifts (sensor noise, unmodeled transients) between the drift-flux simulator and real wells.
[Methods (physics-constraint formulation)] Methods (physics-constraint formulation): The exact mathematical form of the soft mass-conservation constraints, their derivation from engineering principles, the weighting coefficients in the multi-task loss, and the enforcement mechanism are insufficiently specified. This detail is required to assess whether the 65% reduction in negative flows is achieved without introducing bias or limiting generalization.
[Optimization surrogate section] Optimization surrogate section: The >1000× speedup claim for 24-dimensional integrity-aware design optimization depends on surrogate accuracy outside the simulated training distribution; additional quantitative comparison of optimized designs against full drift-flux simulations (e.g., constraint violation rates or objective values) is needed to substantiate this application.

minor comments (2)

[Abstract] Abstract: The phrase 'structural mass conservation with soft physics constraints' would benefit from a one-sentence definition or equation reference to improve immediate clarity for readers unfamiliar with the domain.
[Notation and figures] Notation and figures: Ensure consistent symbol usage for flow rates (q_o, q_w, etc.) and that all result figures include explicit captions distinguishing design-aware versus unaware models and with/without physics constraints.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for their constructive feedback on our manuscript. We have carefully considered each major comment and made revisions to enhance the clarity, reproducibility, and substantiation of our claims. Our point-by-point responses are provided below.

read point-by-point responses

Referee: [§4.2 (ManyWells evaluation)] The central 13× VFM error reduction and 65% negative-flow reduction claims lack specification of the precise error metric (MAE/RMSE/etc.), the exact design-unaware baseline architectures, data splits or cross-validation procedure, and any error bars or statistical tests across the 2000 wells. These omissions are load-bearing because they prevent independent verification of whether design awareness and physics constraints deliver the reported gains.

Authors: We agree these details are essential for verification. We have revised §4.2 to specify the precise error metric, the exact design-unaware baseline architectures, the data splits and cross-validation procedure, and to include error bars and statistical tests across the 2000 wells. revision: yes
Referee: [§4.3 (real-data transfer)] Performance on the five Volve producers is reported solely as R² values with no baseline comparisons, no ablations removing FiLM conditioning or the physics constraints, and no additional metrics (e.g., MAE or negative-flow counts). This is load-bearing for the generalization claim, especially given potential distribution shifts (sensor noise, unmodeled transients) between the drift-flux simulator and real wells.

Authors: We acknowledge this limitation in the original presentation. We have revised §4.3 to include baseline comparisons, ablations removing FiLM conditioning and the physics constraints, and additional metrics such as MAE and negative-flow counts to better support the generalization claim. revision: yes
Referee: [Methods (physics-constraint formulation)] The exact mathematical form of the soft mass-conservation constraints, their derivation from engineering principles, the weighting coefficients in the multi-task loss, and the enforcement mechanism are insufficiently specified. This detail is required to assess whether the 65% reduction in negative flows is achieved without introducing bias or limiting generalization.

Authors: We agree that more detail is needed. The revised Methods section now includes the exact mathematical form of the soft mass-conservation constraints, their derivation from engineering principles, the weighting coefficients in the multi-task loss, and the enforcement mechanism. revision: yes
Referee: [Optimization surrogate section] The >1000× speedup claim for 24-dimensional integrity-aware design optimization depends on surrogate accuracy outside the simulated training distribution; additional quantitative comparison of optimized designs against full drift-flux simulations (e.g., constraint violation rates or objective values) is needed to substantiate this application.

Authors: We recognize the need for further validation of the surrogate. We have added quantitative comparisons of optimized designs against full drift-flux simulations, including constraint violation rates and objective values, in the optimization section. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper presents an empirical ML model (FiLM conditioning + multi-task heads + soft mass-conservation penalties) whose performance claims are tied to held-out evaluation on the ManyWells simulated benchmark and separate R² reporting on Volve real wells. No equation reduces a claimed prediction to a fitted parameter by construction, no self-citation supplies a load-bearing uniqueness theorem, and no ansatz is smuggled via prior work. The derivation therefore remains self-contained against external benchmarks rather than tautological.

Axiom & Free-Parameter Ledger

3 free parameters · 2 axioms · 0 invented entities

Only abstract available so specific free parameters and exact formulations are unknown; model relies on standard ML training plus domain assumptions about well physics and design influence.

free parameters (3)

FiLM modulation parameters
Learned parameters that scale and shift features based on design inputs; number and initialization not specified.
Physics constraint weights
Soft constraint strengths for mass conservation likely tuned during training.
Multi-task loss balancing weights
Weights balancing flow rate, pressure, and classification losses.

axioms (2)

domain assumption Well design parameters influence operational flow behavior
Invoked to justify conditioning embeddings on design via FiLM and attention.
domain assumption Mass conservation holds in wellbore flow
Used to derive soft physics constraints for reducing non-physical predictions.

pith-pipeline@v0.9.0 · 5626 in / 1620 out tokens · 97000 ms · 2026-05-08T06:17:45.848923+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

40 extracted references · 4 canonical work pages · 4 internal anchors

[1]

Virtual multiphase flow metering using diverse neural network ensemble and adaptive simulated annealing

Al-Qutami, T.A., Ibrahim, R., Ismail, I., Ishak, M.A., 2018. Virtual multiphase flow metering using diverse neural network ensemble and adaptive simulated annealing. Expert Systems with Applications 93, 304–316

2018
[2]

Evaluation of commercially available virtual flow me- ters

Amin, A., 2015. Evaluation of commercially available virtual flow me- ters. Offshore Technology Conference, OTC-25764-MS

2015
[3]

Bot- tomhole pressure estimation using evolved neural networks

Ashena, R., Rabiei, M., Rasouli, V., Mohammadi, A.H., 2021. Bot- tomhole pressure estimation using evolved neural networks. Journal of Petroleum Science and Engineering 196, 108013

2021
[4]

Layer Normalization

Ba, J.L., Kiros, J.R., Hinton, G.E., 2016. Layer normalization. arXiv preprint arXiv:1607.06450

work page internal anchor Pith review arXiv 2016
[5]

Multiphase flow in wells

Bai, Y., Bai, Q., Wang, X., 2022. Multiphase flow in wells. In: Subsea Engineering Handbook, 2nd ed. Gulf Professional Publishing

2022
[6]

Physics-informed neural networks for well test interpretation

Bao, J., Li, H., Gao, X., 2022. Physics-informed neural networks for well test interpretation. Journal of Petroleum Science and Engineering 208, 109644

2022
[7]

Oil production monitoring us- ing gradient boosting machine learning algorithm

Bikmukhametov, T., Jaschke, J., 2019. Oil production monitoring us- ing gradient boosting machine learning algorithm. IFAC-PapersOnLine 52(1), 514–519. 41

2019
[8]

First principles and machine learning virtual flow metering: A literature review

Bikmukhametov, T., Jaschke, J., 2020. First principles and machine learning virtual flow metering: A literature review. Journal of Petroleum Science and Engineering 184, 106487

2020
[9]

Modulating early visual processing by language

De Vries, H., Strub, F., Mary, J., Larochelle, H., Pietquin, O., Courville, A.C., 2017. Modulating early visual processing by language. In: Ad- vances in Neural Information Processing Systems (NeurIPS)

2017
[10]

A learned representation for artistic style

Dumoulin, V., Shlens, J., Kudlur, M., 2018. A learned representation for artistic style. In: International Conference on Learning Representations (ICLR)

2018
[11]

A simple approach to virtual flow metering us- ing hybrid mechanistic-data-driven models

Grimstad, B., Gunnerud, V., Sandnes, A.J., Sharika, S., Therrien, J., Ungredda, J., 2021. A simple approach to virtual flow metering us- ing hybrid mechanistic-data-driven models. IFAC-PapersOnLine 54(3), 490–495

2021
[12]

ManyWells: Simulation of multiphase flow in thousands of wells

Grimstad, B., et al., 2026. ManyWells: Simulation of multiphase flow in thousands of wells. Geoenergy Science and Engineering (in press)

2026
[13]

Gaussian Error Linear Units (GELUs)

Hendrycks, D., Gimpel, K., 2016. Gaussian error linear units (GELUs). arXiv preprint arXiv:1606.08415

work page internal anchor Pith review arXiv 2016
[14]

Developing a hy- brid data-driven, mechanistic virtual flow meter—a case study

Hotvedt, M., Grimstad, B., Imsland, L., 2020. Developing a hy- brid data-driven, mechanistic virtual flow meter—a case study. IFAC- PapersOnLine 53(2), 11692–11697

2020
[15]

On the integration of physics-based and data-driven models for virtual flow metering

Hotvedt, M., Grimstad, B., Imsland, L., 2024. On the integration of physics-based and data-driven models for virtual flow metering. Applied Soft Computing, 111168

2024
[16]

Physics-informed machine learning

Karniadakis, G.E., Kevrekidis, I.G., Lu, L., Perdikaris, P., Wang, S., Yang, L., 2021. Physics-informed machine learning. Nature Reviews Physics 3(6), 422–440

2021
[17]

Characterizing possible failure modes in physics-informed neural networks

Krishnapriyan, A.S., Gholami, A., Zhe, S., Kirby, R.M., Mahoney, M.W., 2021. Characterizing possible failure modes in physics-informed neural networks. In: Advances in Neural Information Processing Sys- tems (NeurIPS). 42

2021
[18]

Focal loss for dense object detection

Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P., 2017. Focal loss for dense object detection. In: IEEE International Conference on Computer Vision (ICCV)

2017
[19]

Decoupled weight decay regularization

Loshchilov, I., Hutter, F., 2019. Decoupled weight decay regularization. In: International Conference on Learning Representations (ICLR)

2019
[20]

A machine learning approach for detecting slugging in vertical gas wells

Mask, G., Wu, X., Ling, K., 2019. A machine learning approach for detecting slugging in vertical gas wells. SPE Western Regional Meeting, SPE-195418-MS

2019
[21]

Multimodal anomaly detection for batch distillation using temporal convolutional networks

Nogueira, I.B.R., et al., 2024. Multimodal anomaly detection for batch distillation using temporal convolutional networks. Computers & Chem- ical Engineering, 108830

2024
[22]

Identification of gas-liquid flow regimes using a non-intrusive Doppler ultrasonic sensor and virtual flow regime maps

Nnabuife, S.G., Pilario, K.E.S., Lao, L., Cao, Y., Shafiee, M., 2019. Identification of gas-liquid flow regimes using a non-intrusive Doppler ultrasonic sensor and virtual flow regime maps. Flow Measurement and Instrumentation 68, 101568

2019
[23]

FiLM: Visual reasoning with a general conditioning layer

Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A., 2018. FiLM: Visual reasoning with a general conditioning layer. In: AAAI Conference on Artificial Intelligence

2018
[24]

Physics-informed neural networks: A deep learning framework for solving forward and in- verse problems involving nonlinear partial differential equations

Raissi, M., Perdikaris, P., Karniadakis, G.E., 2019. Physics-informed neural networks: A deep learning framework for solving forward and in- verse problems involving nonlinear partial differential equations. Journal of Computational Physics 378, 686–707

2019
[25]

Neural conservation laws: A divergence-free perspective

Richter-Powell, J., Lorraine, J., Amos, B., 2022. Neural conservation laws: A divergence-free perspective. In: Advances in Neural Information Processing Systems (NeurIPS)

2022
[26]

Drift- flux parameters for three-phase steady-state flow in wellbores

Shi, H., Holmes, J.A., Diaz, L.R., Durlofsky, L.J., Aziz, K., 2005. Drift- flux parameters for three-phase steady-state flow in wellbores. SPE Journal 10(2), 130–137

2005
[27]

Physics-informed machine learn- ing for production forecasting

Zhong, Z., Sun, A.Y., Jeong, H., 2023. Physics-informed machine learn- ing for production forecasting. SPE Journal 28(5), 2700–2717. 43

2023
[28]

Volve field data (2008–2016): Well, reservoir, and pro- duction data from the North Sea

Equinor, 2018. Volve field data (2008–2016): Well, reservoir, and pro- duction data from the North Sea. Available at:https://www.equinor. com/energy/volve-data-sharing. Equinor Open Data Licence

2018
[29]

Norne field reservoir simulation model

OPM Project, 2015. Norne field reservoir simulation model. Available at:https://github.com/OPM/opm-data/tree/master/norne. Open Database License

2015
[30]

Mechanistic Modeling of Gas-Liquid Two-Phase Flow in Pipes

Shoham, O., 2006. Mechanistic Modeling of Gas-Liquid Two-Phase Flow in Pipes. Society of Petroleum Engineers, Richardson, TX

2006
[31]

The elimination of severe slugging—experiments and modeling

Jansen, F.E., Shoham, O., Taitel, Y., 1996. The elimination of severe slugging—experiments and modeling. International Journal of Multi- phase Flow 22(6), 1055–1072

1996
[32]

A multi- task learning approach to virtual flow metering and bottomhole pressure estimation

Sandnes, M., Hotvedt, M., Grimstad, B., Imsland, L., 2021. A multi- task learning approach to virtual flow metering and bottomhole pressure estimation. Knowledge-Based Systems 232, 107458

2021
[33]

When is gray-box model- ing advantageous for virtual flow metering? IFAC-PapersOnLine 55(7), 520–525

Hotvedt, M., Grimstad, B., Imsland, L., 2022. When is gray-box model- ing advantageous for virtual flow metering? IFAC-PapersOnLine 55(7), 520–525

2022
[34]

Inte- grating scientific knowledge with machine learning for engineering and environmental systems

Willard, J., Jia, X., Xu, S., Steinbach, M., Kumar, V., 2022. Inte- grating scientific knowledge with machine learning for engineering and environmental systems. ACM Computing Surveys 55(4), 66:1–66:37

2022
[35]

Scientific machine learning through physics-informed neural networks: where we are and what’s next

Cuomo, S., Di Cola, V.S., Giampaolo, F., Rozza, G., Raissi, M., Pic- cialli, F., 2022. Scientific machine learning through physics-informed neural networks: where we are and what’s next. Journal of Scientific Computing 92, 88

2022
[36]

An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling

Bai, S., Kolter, J.Z., Koltun, V., 2018. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv:1803.01271

work page internal anchor Pith review arXiv 2018
[37]

Attention is all you need

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I., 2017. Attention is all you need. In: Ad- vances in Neural Information Processing Systems (NeurIPS), pp. 5998– 6008. 44

2017
[38]

Modellingflowpatterntransi- tions for steady upward gas-liquid flow in vertical tubes

Taitel, Y., Bornea, D., Dukler, A.E., 1980. Modellingflowpatterntransi- tions for steady upward gas-liquid flow in vertical tubes. AIChE Journal 26(3), 345–354

1980
[39]

On the Opportunities and Risks of Foundation Models

Bommasani, R., Hudson, D.A., Aditi, E., et al., 2021. On the opportu- nities and risks of foundation models. arXiv:2108.07258

work page internal anchor Pith review arXiv 2021
[40]

Stability of stratified gas-liquid flow

Shoham, O., Taitel, Y., 1982. Stability of stratified gas-liquid flow. International Journal of Multiphase Flow 8(3), 217–225. Appendix A. ManyWells Dataset Feature Description Table A.7 lists the operational inputs and well design parameters used in WISE-FM. Appendix B. Hyperparameter Settings 45 Table A.7: Feature descriptions for the ManyWells dataset....

1982