Poly-Directional Scaling and a Recursive Mindset
In most bioprocess development workflows, scale is treated as a linear sequence; bench to pilot to production, where each step serves to stabilize and de-risk the next before advancing. The assumption is that if fundamentals are resolved at small scale, scaling becomes primarily an engineering exercise in maintaining those fundamentals across volume while linearly derisking scale-dependent effects. For many bioprocesses, particularly well-mixed submerged systems with established kinetic models, this approach works reasonably well. The organism's relationship to its environment remains relatively consistent: dissolved oxygen can be controlled, mixing maintained, and gradients minimized through engineering. This deterministic framing becomes somewhat untenable when working with mycelium in solid-state fermentation, where scale-dependent behavior is a defining biological feature of the system itself. Scale-dependent behavior exists in any biological system, of course, but whole-thallus cultivation in solid-state brings a unique and compounding combination of physical, biological, economic, and operational considerations that resist linear unidirectional resolution.
The organism does not experience a 10-liter packed bed the same way it experiences a 100-milliliter flask. Temperature gradients, airflow patterns, moisture stratification, substrate compaction, and the geometry of distance-to-surface all change as volume increases. Mycelium, being physically plastic and holding tremendous capacity for responsiveness, interprets each of these changes with its full phenotypic vocabulary. What appears as stable morphology in small vessels may differentiate into spatial gradients in larger ones. What seems like straightforward colonization kinetics may become conditional on position within the bed, proximity to aeration points, or local thermal conditions (Mitchell et al., 2006).
Critically, these responses are not "scale effects" to be corrected, but reflect the organism reading its environment and adjusting accordingly. The fungus has not become unpredictable; the problem space has expanded, and with it, the organism's opportunity to express context-dependent plasticity. The linear model assumes insights flow in one direction: that bench-scale resolution predicts pilot behavior, which predicts production performance. But when scale itself is a biological parameter this unidirectional flow breaks down. You cannot fully resolve the bench scale without knowing what pilot will reveal, and you cannot interpret pilot behavior without the controlled clarity that bench-scale systems provide.
Under these conditions, it becomes valuable to treat scaling as a multidimensional and multidirectional design space, where scaling up, scaling down, scaling laterally, and scaling digitally occur in parallel as part of a coordinated, adaptive, and evolving learning system. Rather than moving through scales sequentially, this approach moves across them recursively, treating each scale not as a stagegate to pass but as a vantage point that reveals distinct truths about the organism and the process.
Poly-directional scaling
I think of poly-directional scaling as a development strategy that treats scale as a multidimensional, multidirectional design space rather than a linear path. It involves scaling up, scaling down, scaling laterally, and integrating digital modeling in parallel, allowing for recursive learning across distinct system formats, feature targets, and scale-dependent effects. In this framework, scale represents an ensemble of concurrent dimensions, each expressing different physical, biological, and operational truths. Scaling up reveals gradients and emergent behaviors; scaling down sharpens signal and isolates causality; lateral scaling tests format-specific constraints; digital scaling exposes the structure of the response surface itself illuminating hypothetical edge cases and spaces of uncertainty and opportunity. Poly-directional scaling, then, is the practice of resolving these dimensions in concert, using movement across scales as a coordinated learning system rather than a sequence of gates.
Scale Up ⇄ Down
Bi-directional scaling refers to the intentional practice of scaling both up and down in parallel during process and product development. This might mean rapidly advancing a minimum viable process to pilot scale while simultaneously developing medium- or high-throughput bench-scale systems designed to isolate specific variables or mechanisms. The value of this approach lies in its ability to engage with scale-dependent phenomena early in the development cycle while still investing in the clarity and control afforded by small-scale experimentation. Rather than waiting for scale to reveal new feature space after resolution of a foundational feature space, bi-directional scaling enables a recursive exchange of information: insights from the bench refine how we interpret pilot behavior, and emergent behaviors at pilot expose which variables, interactions, or regimes require more disciplined resolution at the bench.
This approach reframes scale as a set of mutually constraining vantage points. Each scale reveals different expressions of fungal plasticity, different operational sensitivities, and different forms of drift. The practice of deliberately moving between scales to surface these differences has precedent in bioprocess engineering, where scale-down models are used to understand gradient effects and heterogeneities that emerge at production scale (Lara et al., 2006; Noorman, 2011). By moving up and down deliberately, and continuously, development teams can detect scale-dependent uncertainties earlier, design scale-dependent features into bench-scale systems, differentiate between artifacts and inherent behaviors, and resolve fundamentals early and, critically, prior to full commitment to scaling infrastructure and assumptions. Bi-directional scaling, in this sense, is less about moving between scales and more about maintaining a dynamic dialogue across them.
Target ⇄ Format
Lateral scaling refers to partitioning distinct response targets across tailored experimental systems to accelerate multi-objective learning. In most development scenarios, and particularly in mycelium, the goal is not to optimize a single outcome but to navigate a constellation of interconnected targets that differ in complexity, responsiveness, and cost to interrogate. Forcing all targets through a single experimental format collapses their differences and slows learning; lateral scaling, by contrast, aligns each target with a system designed to resolve it.
A lower-cost, high-throughput format may be sufficient for a simple or well-behaved target, enabling rapid iteration and statistical clarity. More complex or fragile targets may require bespoke systems that offer higher resolution or more faithful environmental control, even if throughput decreases. The point is not to treat one format as primary and the others as ancillary, but to allow learning to proceed in parallel across systems whose architectures reflect the actual demands of the target. When coupled with adaptive learning tools, lateral scaling creates a network of mutually informing experiments. Insights gained from the faster, cheaper target can shape the design of experiments for the slower, more expensive target. Conversely, mechanistic signals observed in the high-resolution system can refine what the high-throughput system should screen for. This interplay shortens the number of total experiments required to reach a multi-target goal by distributing the cognitive and experimental load across specialized platforms.
Critically, lateral scaling requires designing systems for specificity; matching system to target to question. This specificity, once the exclusive domain of well-funded engineering teams, is now increasingly achievable by individual researchers equipped with contemporary tools. The confluence of open-source hardware platforms and affordable digital fabrication has lowered the barrier to building bespoke experimental systems (Baden et al., 2015; Pearce, 2012). Open source sensor and pump drivers provide modular building blocks for custom bioreactor designs at a fraction of commercial cost. Parametric CAD software and open-source design repositories now allow researchers to iterate rapidly on vessel geometries, sensor housings, gas-distribution components, and other experimental hardware without relying entirely on commercial systems. Consumer-grade 3D printers and CNC mills turn those designs into physical components within hours. Online repositories host validated designs for everything from peristaltic pump controllers to automated humidity chambers, allowing researchers to adapt proven architectures rather than start from scratch.
A reasonably motivated researcher can now envision a system tailored to a specific question, adapt open-source templates, fabricate components in-house or through accessible makerspaces, and develop a functioning platform far more quickly than would be possible through fully commercial or outsourced routes. What once required dedicated engineering support and substantial capital can now often be prototyped more cheaply and quickly by individual researchers or small teams.
This democratization makes format diversification practically feasible. Rather than forcing all targets through a single commercial system, researchers can maintain a constellation of purpose-built platforms, each optimized for specific questions and targets. A high-throughput plate reader might handle one target, while a custom miniature column bioreactor handles another, and a small-scale production geometry analog (perhaps fabricated to preserve scale-relevant gradients) handles a third. Complex, interacting objectives can thus be resolved more coherently and quickly than any single format would allow through bespoke system design within reach of anyone willing to learn.
Physical ⇄ Digital
Adaptive learning refers to a recursive, machine-learning–driven process in which each round of experimentation is designed to reduce uncertainty in the system. Rather than committing to a fixed experimental plan, adaptive learning uses model training and feedback to determine which experimental runs are likely to be most informative given the current state of knowledge. This approach builds on foundational work in sequential experimental design, where experiments are selected iteratively based on accumulating information rather than planned in advance (Fedorov, 1972). Modern implementations integrate Bayesian optimization and active learning to navigate high-dimensional response surfaces, using probabilistic models to balance exploration of uncertain regions against exploitation of promising ones (Shahriari et al., 2016).
At each iteration, the model identifies regions of high uncertainty or high expected value and proposes the next set of experiments accordingly. This keeps experimentation aligned with the primary target as the model evolves and focuses physical effort where learning density is highest. In practice, adaptive approaches can reduce experimental workload by selecting experiments sequentially according to model uncertainty and expected value, although the magnitude of improvement is system-dependent.
In practical terms, experiments are chosen not merely to validate hypotheses, but to actively shape the model's understanding of the system in service of the goal; allowing convergence on outcomes of high practical value while avoiding wasted effort in regions of diminishing return or those which have been reasonably resolved. By explicitly using the evolution of model uncertainty as a function of experiment, adaptive learning reframes progress: the pace of insight, rather than the pace of execution, becomes the meaningful measure.
The digital dimension of poly-directional scaling emerges through the integration of adaptive learning across all physical systems and scales. As experimental data accumulates from bench, pilot, and lateral formats, models can be trained to reconcile both shared and system-specific behaviors. This often requires multiple, co-evolving models: format-specific models that resolve discrete learnings within individual systems, and cross-system models that capture global relationships and scale-dependent effects. Together, these models allow experiments to be designed in coordination, where the next best experiment in one system is informed by what has already been learned elsewhere.
As model resolution improves, it becomes possible to distinguish which features are genuinely scale sensitive and which remain robust, allowing scale-dependent risks to be surfaced and addressed earlier in development. Over time, more exploration can be offloaded to in silico search, accelerating the identification of promising regimes without corresponding physical cost. Throughout this process, the evolution of uncertainty acts as a real-time diagnostic of learning rate, helping teams qualify progress, avoid futile searches, and direct resources toward questions most likely to expand the system's understanding.
Scaling as a Learning System
Taken together, poly-directional scaling is a practical and adaptive framework for navigating the complexity inherent to mycelium R&D. It creates space for early engagement with scale-dependent behavior, distributes learning across multiple tailored systems, and aligns experimentation with the structure of uncertainty rather than a fixed sequence.
Working with fungi across scales requires an almost obsessive habit of deconstructing features and relationships: within individual scale steps, across scale steps, and back again. A density gradient observed at pilot prompts questions about behavior at the bench. An anomaly in bench-scale kinetics becomes reinterpreted after seeing how the organism navigates at intermediate scale. A production-scale failure sends you back to fundamentals, not to repeat what was done before, but to ask what the larger system revealed that the smaller one couldn't. This constant movement (a refusal to treat any scale step as fully resolved independently) is the work. The fungus distributes its logic across the full dimensionality of its response space, and the only way to access it is through persistent, recursive interrogation. In practice, this means resisting the pressure to "lock in" conclusions prematurely, while maintaining an active skepticism about what you think you know at one scale until challenging it at another. It means building systems that allow you to move fluidly between scales rather than treating each as a separate silo. And, perhaps most importantly and most uncomfortably, it means cultivating a tolerance for ambiguity and incompleteness, recognizing that clarity often emerges not from resolving a single question definitively but from holding multiple partial answers in dialogue until their relationships become visible. And most critically, the actual and practical function becomes demonstrable and replicable.
This recursive mindset aligns naturally with the organism itself. Mycelium does not grow linearly; it branches, fuses, reallocates, and reorganizes. It revisits space it has already occupied and adjusts behavior based on new information. To design processes for such an organism is to mirror that recursiveness, and to treat development as a continuous reexamination of relationships, constraints, and possibilities as they unfold across scale.
References
Baden, T., Chagas, A. M., Gage, G. J., Marzullo, T. C., Prieto-Godino, L. L., & Euler, T. (2015). Open labware: 3-D printing your own lab equipment. PLOS Biology, 13(3), Article e1002086. https://doi.org/10.1371/journal.pbio.1002086
Fedorov, V. V. (1972). Theory of optimal experiments (W. J. Studden & E. M. Klimko, Trans. & Eds.). Academic Press.
Lara, A. R., Galindo, E., Ramírez, O. T., & Palomares, L. A. (2006). Living with heterogeneities in bioreactors: Understanding the effects of environmental gradients on cells. Molecular Biotechnology, 34(3), 355–381. https://doi.org/10.1385/MB:34:3:355
Mitchell, D. A., Berovič, M., & Krieger, N. (Eds.). (2006). Solid-state fermentation bioreactors: Fundamentals of design and operation. Springer. https://doi.org/10.1007/3-540-31286-2
Noorman, H. (2011). An industrial perspective on bioreactor scale-down: What we can learn from combined large-scale bioprocess and model fluid studies. Biotechnology Journal, 6(8), 934–943. https://doi.org/10.1002/biot.201000406
Pearce, J. M. (2012). Building research equipment with free, open-source hardware. Science, 337(6100), 1303–1304. https://doi.org/10.1126/science.1228183
Shahriari, B., Swersky, K., Wang, Z., Adams, R. P., & de Freitas, N. (2016). Taking the human out of the loop: A review of Bayesian optimization. Proceedings of the IEEE, 104(1), 148–175. https://doi.org/10.1109/JPROC.2015.2494218