A Simple Sensor Doubled Algae Productivity — And It Could Transform Green Biofuel Farming

Grow too much algae and you end up with less of it. That paradox sits at the heart of a deceptively elegant experiment conducted in the sun-soaked fields of Almería, Spain — and the results hint at what large-scale microalgae farming could look like in the decades ahead.

In a pair of 80-square-metre outdoor ponds, each holding 12,000 litres of swirling green culture, engineers from the University of Almería and the University of Bologna tested a simple but powerful idea. What if, instead of harvesting algae on a fixed schedule, you let a sensor decide? And what if that sensor could tell you — in real time, every five minutes — exactly when the culture was dense enough to harvest? (González-Hernández et al., 2026)

When the researchers switched their automated pond from a target density of 1.0 grams per litre to just 0.8 grams per litre, harvested productivity nearly tripled — jumping from 9.52 to 23.20 grams of algae per square metre per day. The conventional, schedule-driven pond running alongside it managed just 11.16 g m⁻² d⁻¹ over the same campaign. The numbers are striking, but the logic behind them is even more interesting.

Harvested Areal Productivity: Turbidostat vs. Chemostat

Label	Value
RW5 Turbidostat @ 1.0 g/L	9.52
RW5 Turbidostat @ 0.8 g/L	23.2
RW5 Overall Net	20.34
RW6 Chemostat Overall Net	11.16

Source: González-Hernández et al. (2026). Turbidostat productivity shown for each operating phase; chemostat is overall campaign figure.

The Science

Microalgae — single-celled photosynthetic organisms that turn sunlight and CO₂ into biomass — are genuinely promising for a wide range of applications: biofuels, protein-rich animal feed, wastewater treatment, and carbon capture. But industrial-scale production has stubbornly resisted the kind of precision that laboratory systems enjoy. Outdoor ponds are noisy, unpredictable environments, buffeted by clouds, temperature swings, and the relentless cycle of day and night. Keeping a culture at its productive sweet spot has typically relied on operator intuition and fixed daily schedules rather than responsive automation.

The team's solution was a turbidostat — a term worth unpacking. In a standard chemostat, a bioreactor is diluted at a fixed rate regardless of what the culture is actually doing. In a turbidostat, dilution is triggered by the culture's actual density: when it gets too thick, fresh medium flows in and the overflow carries excess algae out. Turbidostats are well-established in laboratory settings, but scaling them to a semi-industrial outdoor raceway — a long, shallow, oval channel where culture is kept circulating by a paddle wheel — had never been demonstrated convincingly before this work.

The outdoor raceway photobioreactor at IFAPA (the Andalusian Institute for Agriculture Research) in Almería is exactly that kind of large-scale channel system, covering 80 m² with culture flowing in a continuous loop

. It is the workhorse of large-scale algae production globally: cheap to build, easy to scale, but difficult to control with precision.

The algae species used was Scenedesmus almeriensis, a robust green microalga well-suited to the Mediterranean climate. Culture depth was held at 15 cm — the optimal operating height established in previous work for this reactor type — giving a working volume of roughly 12,000 litres per pond.

Two geometrically identical ponds were run in parallel throughout the 14-day campaign. RW5 ran the new turbidostat strategy; RW6 ran as a conventional chemostat, receiving a fixed daily dilution of 20% of its total volume (2,400 litres) on operating days. Both ponds experienced identical weather and the same starting biology, providing meaningful context for comparing results.

The optical sensing system at the core of the experiment is worth examining closely, because its low cost and open design are central to the paper's practical ambitions

. Rather than expensive laboratory-grade turbidity meters, the team used a pair of off-the-shelf multispectral sensors (the AS7341 by AMS-OSRAM, a chip costing a few euros) connected to a Raspberry Pi. Culture from the pond was pumped through a small flow-through cuvette every five minutes. Two optical measurements were taken simultaneously: an absorbance measurement, where white light passes through the sample to a detector, and a fluorescence measurement, where a 450 nm blue excitation source causes chlorophyll to emit red light, captured at 90° to filter out direct illumination scatter. Between measurements, the cuvette was flushed with clean water to prevent fouling.

The raw optical signals were converted to biomass concentrations using a LASSO regression model — a form of regularised linear regression ($L_1$ penalty) that automatically selects only the most informative spectral features, preventing the model from overfitting to noise. The resulting estimator, $\hat{X}(t)={\beta }_0+{\sum }_{i=1}^n{\beta }_i{s}_i(t)$, produced a biomass estimate in grams per litre every five minutes throughout the campaign.

The turbidostat control logic itself was refreshingly simple: a hysteresis-based on-off switch. When $\hat{X}(t)$ exceeded the upper threshold $X_{\max }$, the dilution valve opened; when it fell below $X_{\min }$, the valve closed. Critically, dilution was only permitted between 09:00 and 20:00 — daylight hours — preventing the system from wastefully flushing biomass out of the pond at night when growth has stopped. The valve state is formally described as:

$$u(t)=\{\begin{array}{ll}1,& \text{if }\hat{X}(t)>X_{\max }\text{and}t\in {\mathcal{T}}_{\text{light}}\\ 0,& \text{if }\hat{X}(t)<X_{\min }\text{or}t\notin {\mathcal{T}}_{\text{light}}\\ u(t^{-}),& \text{otherwise}\end{array}$$

Commands travelled from a MATLAB supervisory layer to a PLC (programmable logic controller — the industrial standard for reliable on-site automation) via an OPC UA communication link, a widely used industrial protocol. The architecture

was deliberately chosen to be compatible with existing industrial infrastructure, not a bespoke laboratory system.

What They Found

The biomass estimator performed remarkably well for a low-cost system. Evaluated against 18 pairs of offline dry-weight measurements — the gold-standard method for biomass quantification — it achieved a mean absolute error (MAE) of 0.042 g L⁻¹ and a root mean square error (RMSE) of 0.052 g L⁻¹. For reference, the operating thresholds were set at 0.8–1.0 g L⁻¹, so the sensor was accurate to within roughly 5% of the full operating range. That was precise enough for the hysteresis controller to distinguish "harvest now" from "wait" reliably, without chattering.

The turbidostat's ability to hold the pond at a target concentration was clearly demonstrated across the two operating phases

. From 28 April to 4 May, with the threshold set at 1.0 g L⁻¹, RW5 harvested a total of 5.33 kg of biomass at a harvested areal productivity of 9.52 g m⁻² d⁻¹. Then, on 5 May, the threshold was intentionally lowered to 0.8 g L⁻¹. A large transition harvest that day flushed out the excess biomass — that day is excluded from steady-state analysis for good reason. From 6 to 10 May, operating at the lower density, the pond harvested 9.28 kg and reached 23.20 g m⁻² d⁻¹. The jump of 144% in just one parameter change is the paper's headline result.

Daily Harvested Biomass: Turbidostat (RW5) vs. Chemostat (RW6)

Label	Value
Apr 27	10,489
Apr 28	928.32
Apr 29	1,014.9
Apr 30	1,023.33
May 01	534.85
May 02	747.12
May 03	512.37
May 04	572.6

Source: González-Hernández et al. (2026). RW6 shows zero on days when no chemostat dilution was scheduled (weekends).

The overall productivity picture, accounting for all days of the campaign and the full biomass balance (not just harvested output), showed the turbidostat pond at 20.34 g m⁻² d⁻¹ versus the chemostat pond's 11.16 g m⁻² d⁻¹ — an 82% advantage. Even this comparison carries a caveat the authors are careful to state: the two ponds were optimised for different goals, so this isn't a controlled A/B trial. RW6's chemostat operation actually struggled under these conditions — its biomass concentration declined steadily over consecutive harvesting days because the fixed 20% daily dilution was too aggressive for the prevailing growth rates, only recovering during weekend batch periods when no harvesting occurred.

Total Harvested Volume Over Campaign

Label	Value
RW5 Turbidostat	31,330
RW6 Chemostat	21,600

Source: González-Hernández et al. (2026). Each reactor holds 12,000 L total volume.

Why This Changes Things

The self-shading effect that explains the productivity jump deserves a fuller examination, because it challenges a natural intuition. More algae should mean more biomass produced, right? Not quite. Algae need light to grow, and in a dense culture, cells near the surface absorb most of the incoming photons before they can penetrate to cells deeper in the pond. This self-shading reduces the average light available per cell. Beyond a certain density threshold, the culture becomes its own worst enemy. Conversely, a thinner culture allows light to penetrate more deeply, each cell receives more photons on average, and the whole population grows faster. The trick is finding the optimal density — not so thin that you're wasting pond space, not so thick that photons are trapped near the surface. This trade-off is well-documented in the literature (González-Hernández et al., 2026), but it has been devilishly hard to exploit at scale because it requires real-time knowledge of what the culture is actually doing.

That's exactly what this sensing-and-control loop provides. And crucially, it provides it cheaply. The sensing hardware — a Raspberry Pi, two spectral chips, a small pump, and a flow-through cuvette — costs orders of magnitude less than the NIR turbidity probes used in comparable research installations like the AlgaePARC facility studied by de Vree et al. (referenced in González-Hernández et al., 2026). The web interface for real-time visualisation

runs on the same embedded hardware. The whole system was deployed outdoors, in an enclosure next to the raceway, and ran continuously for two weeks without reported failure.

For context on what these productivity numbers mean: global microalgae production today sits at around 10,000–20,000 tonnes per year, largely constrained by the economics of cultivation. A near-doubling of productivity per square metre of pond area, achieved with a control upgrade rather than new hardware, is the kind of improvement that can shift those economics meaningfully. Raceway ponds are already the lowest-cost large-scale cultivation technology; making them smarter rather than bigger is exactly the right direction.

The daylight-only operating constraint is another underappreciated contribution. Previous turbidostat demonstrations have largely been in laboratory settings where illumination is continuous. In outdoor systems, harvesting at night removes biomass that isn't being replaced by photosynthesis — a pure loss. The simple time window ${\mathcal{T}}_{\text{light}}=[09:00,20:00]$ encoded in the control law is a small design choice with significant practical consequences, and it's the kind of insight that only emerges from real outdoor operation rather than simulation.

The experiment also directly demonstrated operator flexibility: the ability to deliberately shift the operating concentration mid-campaign — from 1.0 to 0.8 g L⁻¹ — and have the system smoothly transition to the new setpoint without manual intervention. This is not merely convenient. It is the prerequisite for a more sophisticated layer of optimisation: economic model predictive control (EMPC), a framework in which a higher-level algorithm would forecast weather conditions, solar radiation, and market signals to continuously adjust the optimal biomass target. The turbidostat and sensor presented here are explicitly framed as the "enabling layer" for that future capability (González-Hernández et al., 2026).

What's Next

The 14-day experiment is a strong proof of concept, but it raises as many questions as it answers. The most immediate is whether 0.8 g L⁻¹ is truly optimal for Scenedesmus almeriensis in Almería in late April and early May, or whether an even lower density would yield even higher productivity — and where the floor is. The authors note that there is a trade-off between light utilisation and self-shading, implying an optimal biomass concentration range, but this campaign only sampled two operating points.

Seasonal variation is another open frontier. The experiment ran in spring conditions in southern Spain. Summer months bring higher solar radiation, which may shift the optimal density; winter months bring lower radiation and potentially slower growth rates that could make fixed-schedule chemostat operation comparatively more competitive. A full year of turbidostat operation, tracking the seasonally optimal setpoint, would be the natural next study.

The comparison with the chemostat reactor, while informative, was not designed as a rigorous controlled experiment. The two ponds were not inoculated identically and operated under different biomass trajectories. A properly controlled multi-season comparison — with the chemostat dilution rate optimised for prevailing conditions rather than fixed at 20% — would give a cleaner picture of the true turbidostat advantage.

The sensing system itself, while impressively accurate for its cost, measured biomass at a single point: the flow-through cuvette mounted on the side of the pond. Raceway systems are known to have spatial heterogeneity — biomass concentration can vary along the 40-metre channel length due to localised growth, mixing gradients, and CO₂ injection patterns. Whether a single sensor placement captures the pond-average biomass reliably across all conditions is worth probing more carefully.

Finally, the economic case, though implicit in the productivity numbers, deserves explicit modelling. A techno-economic analysis comparing turbidostat operation (sensor cost, control infrastructure, reduced waste from over-dilution) against conventional chemostat operation at different scales would clarify where this approach makes commercial sense. The authors' framing of this work as infrastructure for future EMPC-based optimisation is compelling — but the pathway from this 14-day trial to a fully autonomous, economically optimised algae farm is not yet mapped.

What González-Hernández et al. (2026) have done is remove one of the most persistent excuses for not trying: the claim that real-time biomass control in large outdoor algae ponds is too complicated or too expensive to bother with. A Raspberry Pi, two spectral chips, a hysteresis controller, and careful attention to daylight constraints produced a near-doubling of productivity. The next step is not another proof of concept — it's deployment at scale, across seasons, species, and latitudes, to find out just how far this simple idea can go.