🔬 physics

PRIMS: Physics-guided Representation for Fluid Identification in Multimodal Sensing

The paper introduces PRIMS, a physics-aware multimodal Transformer that integrates fluid mechanics principles into its architecture to achieve highly accurate, data-efficient, and robust on-device fluid identification under varying operating conditions, significantly outperforming existing state-of-the-art methods while using far fewer parameters.

Original authors: Hai-Long Nguyen, Trung Thanh Nguyen, Lars Holm, Dennis Alveringh, Duc Viet Le

Published 2026-07-27

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Hai-Long Nguyen, Trung Thanh Nguyen, Lars Holm, Dennis Alveringh, Duc Viet Le

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to identify a mystery liquid flowing through a tiny, invisible pipe. You can't see the liquid, and you can't taste it. All you have are two special sensors: one that feels how heavy the liquid is as it rushes past (like a scale for moving water), and another that feels how hard it is to push the liquid through (like feeling the resistance of honey versus water). This is the world of microfluidics, a field where scientists and engineers manipulate tiny amounts of fluids for everything from medical tests to environmental monitoring. The challenge? These liquids behave differently depending on how fast they flow, how hot they are, or how much pressure is pushing them. Traditional computer programs often get confused by these changes, acting like a student who memorized answers for one specific test but fails when the questions are slightly different. They treat the sensor signals as just random numbers, missing the hidden "rules of the road" that physics uses to govern how fluids move.

Enter PRIMS, a new kind of smart computer program designed to solve this puzzle. Think of PRIMS not as a robot that blindly memorizes data, but as a detective who actually understands the laws of physics. While other programs might just look at the raw numbers and guess, PRIMS is built with a built-in "physics brain." It knows that if the liquid gets thicker (more viscous), the pressure changes in a specific way, and if the liquid gets heavier, the vibration of the sensor changes in another. By weaving these physical rules directly into its brain, PRIMS can figure out what the liquid is—even if it's flowing at a speed or temperature it has never seen before. The researchers found that this approach is not only smarter but also much smaller and faster than the giant, complex programs currently used, making it perfect for tiny devices that need to work on the go.

The Problem: When Smart Computers Get Stuck

For years, scientists have tried to use machine learning—computer programs that learn from examples—to identify fluids. Imagine showing a computer a thousand pictures of water, a thousand of oil, and a thousand of alcohol, and asking it to learn the difference. This works well if the conditions stay the same. But in the real world, things change. If the temperature rises, or the liquid flows faster, the sensors send different signals. Standard programs, which treat these signals as just "data," often get confused. They might think a fast-flowing glass of water is actually oil because the numbers look similar, even though the physics are totally different.

The researchers argue that these programs are missing the point. They are ignoring the physical relationships that connect the sensors. For instance, there is a famous rule in physics (called the Hagen–Poiseuille law) that links how fast a fluid moves, how thick it is, and how much pressure it takes to push it through a pipe. Old programs ignore this rule and try to guess the answer purely from patterns in the data. The paper suggests that this is why they fail when conditions change: they haven't learned the rules, they've just memorized the examples.

The Solution: PRIMS, the Physics-Savvy Detective

To fix this, the team created PRIMS (Physics-guided Representation for Fluid Identification). Instead of feeding raw sensor data directly into a giant, generic brain, PRIMS breaks the process down into three clever steps, each designed to respect the laws of physics.

1. The Translator (Physics-based Token Vectorization)
First, PRIMS acts like a translator. Raw sensor signals are messy waves of electricity. PRIMS doesn't just swallow them whole; it translates them into "tokens" that make physical sense. It looks at the Coriolis sensor (the one measuring weight and density) and the pressure sensor, and it asks: "What physical property does this wave represent?" It converts the messy data into clean, meaningful concepts like "mass flow" and "pressure drop." It's like taking a chaotic jumble of musical notes and organizing them into a clear melody that tells a story about the fluid's behavior.

2. The Synthesizer (Physical Component Synthesizer)
Next, PRIMS uses a special module to figure out the fluid's viscosity (how thick or sticky it is). In the real world, viscosity is the link between how fast the fluid moves and how much pressure is needed to push it. PRIMS uses a technique called "cross-attention" to let the "pressure" part of the data talk to the "flow" part. Imagine two detectives sharing clues: one says, "The pressure is high," and the other says, "The flow is slow." Together, they deduce, "Ah, this must be a thick, sticky fluid!" This step ensures the computer isn't just guessing; it's calculating the relationship between the variables just like a physicist would.

3. The Fusion (Physics-guided Fusion)
Finally, PRIMS brings all these clues together. It takes the translated signals and the calculated viscosity and blends them using a "self-attention" mechanism. This allows the model to weigh which clues are most important for identifying the specific fluid. It's like a judge listening to all the evidence before making a final verdict. Because the clues were built on physical laws, the verdict is much more reliable, even if the case is tricky.

The Results: Smarter, Smaller, and Stronger

The researchers tested PRIMS on five different fluids: Nitrogen (a gas), Water, Isopropanol, Ethanol, and Acetone. They put these fluids through a gauntlet of changing conditions, including different temperatures, pressures, and flow speeds.

The results were impressive. PRIMS achieved an average accuracy score (F1-score) of 98.92%. To put that in perspective, the best previous "state-of-the-art" computer models (like iTransformer) got about 97.37%. But the real magic is in the size. The giant models that came close to PRIMS' accuracy were huge, with about 6.5 million parameters (think of these as the number of connections in the computer's brain). PRIMS, however, did it with only 0.46 million parameters. That is a 14 times reduction in size!

This matters because smaller models can run on tiny, battery-powered devices (like the microfluidic chips used in hospitals or environmental sensors) without needing a massive computer server. PRIMS is not just accurate; it's efficient.

Why It Matters: Learning from the Rules, Not Just the Examples

The most exciting part of the paper is how PRIMS handles "out-of-distribution" tests. This is a fancy way of saying: "What happens if we test it on a temperature or flow speed it has never seen before?"

When the researchers tested the models on temperatures they hadn't trained on, the old models started to crumble. For example, at a specific temperature, a competing model might drop its accuracy significantly because it had never seen that specific combination of numbers before. PRIMS, however, stayed strong. Because it learned the underlying physical rules (like how viscosity changes with temperature), it could adapt to new situations without needing to be retrained. The paper suggests that by embedding these physical relationships directly into the architecture, the model learns to be "environment-independent."

The researchers also looked at the "brain" of the model to see what it was learning. They found that the part of PRIMS responsible for "dynamic viscosity" created very clear, distinct groups for each fluid, almost like sorting marbles by color. This confirms that the model wasn't just memorizing; it was actually learning the physical differences between the liquids.

The Bottom Line

PRIMS shows that when you teach a computer the rules of physics, it becomes a much better detective. It doesn't just memorize the past; it understands the present and can predict the future, even in conditions it hasn't seen before. By combining deep learning with the timeless laws of fluid mechanics, the authors have created a tool that is not only more accurate but also small enough to fit on the tiny chips of the future. While the paper notes that it currently focuses on simple, smooth flows (laminar flow) and could be improved by adding even more physical constraints, the results suggest a powerful new direction: building AI that respects the laws of nature.

Technical Summary: PRIMS – Physics-guided Representation for Fluid Identification in Multimodal Sensing

Problem Statement

Accurate on-device fluid identification is critical for microfluidic applications in industrial, biomedical, and environmental sectors. While modern sensors (specifically Coriolis mass-flow and differential pressure sensors) provide rich multimodal data, existing learning-based methods face significant limitations:

Lack of Physical Awareness: Current models often treat sensor signals as domain-agnostic features, neglecting the underlying physical relationships (e.g., between flow rate, pressure, and viscosity) that govern fluid behavior.
Poor Generalization: Models trained on specific conditions often fail when operating parameters (temperature, pressure, flow rate) shift, as they rely on memorizing patterns rather than learning transferable physical principles.
Computational Inefficiency: State-of-the-art Transformer-based models (e.g., iTransformer) capable of handling these time-series tasks are often too large (millions of parameters) for resource-constrained edge devices.

Methodology: PRIMS

The authors propose PRIMS (Physics-guided Representation for Fluid Identification in Multimodal Sensing), a physics-aware multimodal Transformer designed to embed fluid dynamics knowledge directly into the model architecture. The framework processes raw Coriolis ( $C_1, C_2$ ) and pressure ( $P_u, P_d$ ) signals through three dedicated modules:

1. Physics-based Token Vectorization (PTV)

Instead of feeding raw time-series data directly into the network, PTV transforms signals into compact, physically meaningful embeddings:

Coriolis Processing: Applies Short-Time Fourier Transform (STFT) and notch filtering to extract magnitude and phase spectra. These features are mapped to mass flow rate and fluid density based on the physical principles of tube oscillation.
Pressure Processing: Computes mean ( $\mu_P$ ) and differential ( $\Delta P$ ) pressure features to capture steady-state and transient flow characteristics.
Embedding: These features are projected into a shared embedding space with sinusoidal positional encoding to preserve temporal dependencies.

2. Physical Component Synthesizer (PCS)

This module models the intrinsic dependencies among flow, pressure, and density, specifically targeting viscosity estimation based on the Hagen–Poiseuille law ( $\Delta P \propto \nu \cdot \dot{m}$ ).

Kinematic Viscosity Module: Uses bidirectional cross-attention to align Coriolis (flow/density) and pressure representations. Two separate attention blocks ensure both modalities contribute equally to the kinematic viscosity representation, reflecting the physical ratio $\nu \propto \Delta P / \dot{m}$ .
Dynamic Viscosity Module: Derives dynamic viscosity by attending the fused kinematic viscosity embedding to density-related features (from the Coriolis modality), effectively modeling the relationship $\eta = \nu \rho$ .

3. Physics-guided Fusion (PGF)

The PGF module integrates all representations (flow, pressure, density, and viscosity) using multi-head self-attention.

It retains the directional outputs of the PCS modules to allow the model to weigh coupling directions independently per fluid.
It employs an Adaptive Spatial Sync (ASS) module and residual connections to recalibrate channel-wise features, ensuring the final fused representation captures cross-physical dependencies while maintaining interpretability.

A lightweight classification head (Global Average Pooling + FC layers) then predicts the fluid type from the fused representation.

Key Contributions

Architecture Design: Introduction of PRIMS, a compact Transformer that explicitly embeds physical relationships (flow, pressure, density, viscosity) into tokenization and attention layers, enabling efficient deployment on edge devices.
Three-Stage Physics Integration: The design of PTV, PCS, and PGF modules that jointly transform raw signals into physics-consistent representations, synthesize viscosity features, and fuse multimodal data.
Performance and Efficiency: Demonstration that PRIMS achieves superior accuracy with a significantly reduced parameter count compared to existing SOTA models.

Experimental Results

The model was evaluated on a benchmark dataset containing five fluids (Nitrogen, Water, Isopropanol, Ethanol, Acetone) under varying temperatures (288K–308K), pressures (4–6 bar), and flow rates.

Accuracy: PRIMS achieved an average F1-score of 98.92% (±0.03%), outperforming the best baseline, iTransformer (97.37%), and significantly surpassing purely data-driven models like ConvNet and BiLSTM.
Efficiency: PRIMS uses only 0.46 million parameters, representing a 14× reduction compared to iTransformer (6.42M) and a 25× reduction compared to TimeXer (8.53M).
Robustness (Out-of-Distribution): In experiments where entire temperature regimes were held out during training, PRIMS consistently outperformed baselines. For instance, at 298K (unseen during training for that specific split), PRIMS achieved 96.91% average F1, while iTransformer dropped to 78.93% and TQNet to 78.09%.
Ablation: Removing any of the three physics-aware modules resulted in performance degradation, confirming their complementary roles. Specifically, removing the fusion module caused a 2.62% drop, highlighting its necessity for integrating cross-physical relationships.
Latent Space Analysis: PCA visualization showed that the dynamic viscosity branch learned highly discriminative features, separating all five fluid classes clearly, while kinematic branches captured complementary or overlapping characteristics consistent with their physical properties.

Significance and Claims

The paper claims that PRIMS demonstrates that designing architectures which explicitly mirror governing physical relationships allows models to learn transferable, environment-independent representations. By bridging analytical fluid mechanics and deep learning, PRIMS achieves:

Interpretability: The model's internal states correspond to physical quantities (viscosity, density) rather than abstract features.
Resilience: Strong generalization to operating conditions (temperature, flow) not observed during training.
Practicality: High accuracy with a small footprint suitable for embedded microfluidic sensing systems.

The authors note limitations, specifically that the current implementation relies on architectural design rather than explicit auxiliary loss functions to enforce physical constraints, and that the Hagen–Poiseuille assumption (laminar, incompressible flow) may introduce errors for gases at high flow rates. Future work aims to incorporate explicit physical constraint losses and extend the framework to non-Newtonian and compressible flows.