Designing environmentally efficient aquafeeds through the use of multicriteria decision support tools

By: Ronan Cooney, Alex H. L. Wan, Fearghal O’Donncha and Eoghan Clifford *

Aquaculture is the fastest growing food production system, and the sector accounts for more than half of all fish consumed. Its potential as a sustainable food source has been recognised within the EU Farm to Fork strategy and by the targeting of EU Green Deal research funds. Aquafeed is the primary source of environmental cost in farmed finfish and shrimp life cycle assessments (LCA), and thus, emerging ingredients have a key role to play in increasing its sustainability.

High value aquaculture species (e.g., salmon, penaeid shrimp and seabass) are often carnivorous, and as such, require high-protein diets, which traditionally have relied on fish oil and meal. Recent reviews of LCA and circular economy studies in aquaculture have high- lighted a need to develop tools and methods that can enhance its sustainability.

The environmental burden associated with feed is not unique to aquaculture and can be found in pigs, poultry and cattle production systems; however, terrestrial farmed animal feeds are perceived as less environmentally challenging than aquafeeds due to the lower protein requirement (i.e. omnivorous/herbivorous vs carnivorous fish species).

Feed use in aquaculture not only accounts for the bulk of environmental impact but also 40%e70% of monetary production costs. Thus, significant efforts have been made to source alternative, lower-cost ingredients. Research has focused on the replacement of fish meal with cheaper and potentially less environmentally burdensome ingredients, such as plant by-products, algae (micro and macro), insects, land animal by- products and single-cell proteins (including bacteria and yeast).

Alternative protein sources need to meet a number of criteria if they are to be deemed commercially viable. These criteria include (i) nutritionally adequate (e.g. digestible and does not significantly impair fish growth performance or physiological and health status), (ii) palatable to the fish, (iii) scalable to commercial levels, (iv) physically stable, (v) easily handled and stored and (vi) crucially, have lower environmental and life cycle impacts. Given the complex criteria that feeds are required to meet, there is a gap in the tools available for feed formulation that balance economic and environmental efficiencies.

This paper proposes a multicriteria decision support tool that leverages machine-learning techniques and presents a conceptual framework to interrogate these disparate data- sets to identify more efficient feed formulations faster.

Towards sustainable feeds—replacing trial and error approaches

Advanced computational and statistical modelling, through machine learning, have been applied to areas such as medicine, renewable energy and wastewater treatment to optimise the design or operation of various products and processes. Such a step does not preclude the need for in vivo or in vitro experimental validation but enables a more targeted approach to the initial selection of product features.

The use of machine-learning capabilities in aquaculture has previously been realised in disease detection, water quality monitoring, on-site management, farm site selection, risk assessment and in feeding regimes. Machine-learning applications in aquaculture feed formulation (and indeed other food sectors reliant on complimentary feed) can provide a bridge between disciplines such as industrial ecology, biology, environmental engineering and nutrition to arrive at precise aquafeeds.

In one conceptual example, the use of neural networks could allow for the use of meta-datasets and literature to help design options for a least-cost feed formulation, which is species-specific, using low environ- mental impact ingredients, while shortening the supply chain and optimising fish growth. The use of machine learning as part of this formulation allows for a bolt-on approach to these established methods and may present new possibilities with regards to automation and process efficiency while aligning with emerging policies.

By iterating new product prototypes through ‘virtual’ trial and error experiments that also simulate fish response, the feed development paradigm can be revolutionised. Emerging trends of explainable machine learning can further be used to interrogate meaningful connections within feed formulations and associated fish response for enhanced insight.

Data availability

The links between aquafeed ingredients, nutritional profile and environmental impact have been included in several major feed ingredient studies and databases. Environmental training data can be sourced from life cycle inventories such as the ECOALIM database. The ECOALIM database was designed as a tool, which could be used by feed manufacturers and LCA practitioners to investigate the environmental impacts of a particular aquafeed formulation.

The database is a landmark in the development of open-access information concerning feed formulation, impacts and ingredients. In lieu of experimental data, public and proprietary datasets can be used to develop machine learning models. These systems collate feed data and act as repositories for information on established, historical and emerging feed ingredients. Life cycle databases such as ecoinvent and Agri-footprint can be integrated with the afore mentioned nutritional datasets. Economic data on feed ingredients could be sourced from online databases, price indexes or through industry partnerships.

Feed to farm

Extracting relevant data for each of the ingredients from the nutritional, economic and life cycle datasets can allow for the first step in the development of a broader feed design tool. The next step involves transforming the data into a form amenable for use in specific machine learning algorithms. Data are aligned into a pair of row-aligned matrices termed features or inputs and labels or responses, from which the model learns the complex mapping between these paired datasets.

This step also considers data preprocessing, such as treatment of data gaps, outliers or uncertain data and various range normalisation steps that might be acted on the data to eliminate spurious variations that can skew the learning algorithm. Some of the actions done as part of this feature selection process can include transformations guided by external models and physics based approaches that combine multiple datasets into distinct features.

Feature engineering is generally considered the most important consideration to ensure a successful machine learning model. The follow-on step to this is to train the model and evaluate its results. A core part of this model training is adjusting model hyperparameters to best represent the data. A number of libraries exist to simplify these hyperparameter optimisation steps, such as the Lale semi-automated machine learning framework. Following this, the model generates a prediction that aims to meet nutritional requirements, optimise costs and reduce environmental impact.

Using such computational programming approaches, one can uncover hidden patterns in complex environmental, economic and nutritional datasets that can be used to generate efficient feed formulations faster. This can improve the efficiency of trial-and-error approaches (which it should be said are backed up by significant expertise) and in turn, generate more optimised feeds that can be tested and embedded into sustainable feed supply chains. This switch from a model-centric to a data-centric approach means that there is no restriction on the type of data that is to be considered and enables nonlinear relationships to be interrogated.

The development of a multicriteria decision support tool is a step to optimise economic costs, maintain the nutritional value and nutrient digestibility while minimising the environmental impacts. Assessing these factors rapidly in a simulated manner can enable feed manufacturers and researchers to disrupt the feed development process, optimise the design of feeds before the trial stage, and target the design of feeds with specific priorities in mind. This tool should be grounded in life cycle thinking.

As such, it is necessary to expand the boundaries of the tool and connect it with other previously proposed machine learning-based applications and tools. Such an expansion would allow for the accounting of direct, indirect and induced impacts, which feed production and use, can have on the life cycle of aquaculture products. Efficient feed formulation can only increase the sustainability of aquaculture so far.

A finalised or commercial version of the discussed approach could include the use of blockchain to increase traceability and security through the supply chain while meeting consumer or market demands.

Blockchain has in recent years become a popular means of ensuring the providence of ingredients and products. Its inclusion and use in aquafeeds can help to bolster consumer confidence and perceptions in the face of negative publicity, such as those surrounding supply chains of fish meal. With more food products and production systems using blockchain as a means of increasing traceability, its inclusion into the aquafeed production chain can position aquaculture as a prime example of a resource-efficient and traceable food production system. This combination of efficient feed and advanced traceability of the final farmed product could appeal to environmentally conscious consumers.

The adopting of a data-centric rather than model-centric approach introduces a paradigm shift where the large volumes of (training) data available allow a machine learning model to learn the pertinent (nonlinear) relationships and mappings between given inputs and outputs.

“This approach has enormous advantages in terms of simplifying deployments and allowing models to be applied to a wide variety of feed formulations (since only the data changes).”

By using a data-centric approach, there are no restrictions to data that can be considered. The application of machine learning to the questions posed by sustainable feed formulation opens up the field to a technology that has made enormous strides across multiple industries over the past decade. This article outlined and presented the benefits for the interlinking of economic, nutritional and environmental impact datasets to develop an innovative framework for feed formulation design for farmed aquatic animals using machine learning.

There are numerous datasets and databases available to develop the structure and tools necessary to implement such a system, the bene- fits that this avenue of research has to offer can increase the sustainability of aquaculture and strengthen its role in feeding the world.

This is a summarized version developed by the editorial team of Aquaculture Magazine based on the review article titled “Designing environmentally efficient aquafeeds through the use of multicriteria decision support tools” developed by: Ronan Cooney, Alex H.L.Wan, Fearghal O’Donncha and EoghanClifford. The original article was published in May 2021 via Science Direct.

Leave a comment

Your email address will not be published. Required fields are marked *