DTs are virtual representations of living or non-living physical entities. Enabled by improvements in computing capabilities, they exist in silico, that is, as computer-simulated models4. Deployment of sensors that detect biological, chemical, and physical properties of objects in real-time, ensures that the digital counterparts of these measured objects are accurate and ‘live’5. In such cyber-physical architectures, changes that occur in the physical system are modifying its virtual twin simultaneously and continuously.
With origin in experimental designs of satellites, spacecrafts, city infrastructures6and civil engineering writ large, in recent years, DTs have been re-purposed to address predicaments such as climate change and extreme weather, in complex natural environments7.8.
By simulating the state of physical systems, DTs can be queried using advanced modeling techniques to uncover optimal behavior. Reinforcement Learning (RL), a subfield of AI that enables autonomous agents to make decisions in complex systems9, can be deployed in DTs to advise optimal control strategies to the physical counterpart. RL agents take the current state of a system as input, and predict future action sequences that optimize system behavior. DTs allow agents to simulate many control sequences to determine which aligns best with the control objective before advising the physical system.
Combining virtual replicas with such advanced decision-making technologies will have profound transformative implications for the agrifood sector10, offering possible remedies to the problems of malnutrition, GHG emissions, and food waste. To appreciate these prospects, we acknowledge potential applications across six supply chain steps: (a) agricultural inputs, (b) primary agricultural production, (c) storage and transportation, (d) food processing, (e) distribution and retail, and ( f) consumption (Fig. 1).
Inputs for agricultural production
Agricultural inputs commonly refer to agro-chemicals, such as nitrogen (N) and phosphorous (P) fertilizers, pesticides, and crop seeds, which are essential for yield productivity. The carbon footprint involved in the manufacture of these inputs is considerable. For example, CO2 emissions of N fertilizer production in China is estimated at 452 Tg CO2-eq, constituting 7% of total GHG emissions from the Chinese economy. Measures to improve heat conversion efficiency in power plants supporting N fertilizer manufacturing are recognized as an essential intervention to lower carbon intensity11th. In this context, ‘virtual power plants’ could be developed and used by RL agents to find control policies that maximize electricity generation whilst minimizing CO2 emissions12.
DTs proven to operate at the molecular, cell, tissue and organ levels5 can enable precise simulations of crops. New ‘virtual crops’ could be rapidly stress-tested in computer laboratories under alternate conditions, including precipitation, temperature and salinity, to discover desirable traits and risk factors. While genetically modified organisms (GMO) are currently precluded in some jurisdictions, including the European Union, in the face of a shifting climate niche13 such laboratories could prove useful in supporting seed improvements for climate-resilient staples.
primary agricultural production
Beyond the organ level, virtualization of entire farming systems that replicate atmospheric factors, geomorphological processes and edaphic conditions, including soil microbiology, would support precision agriculture at unprecedented scales. Such DTs are likely to use cameras and sensors to sample humidity, moisture content, temperature, irradiance, irrigation and nutrient supply as often as every minute. The digitalisation of agricultural production has the potential to revolutionise problems in animal health, farming resource efficiency and biodiversity loss14,15,16.
DTs can be used to actively monitor livestock well-being using facial recognition technology that infers emotion from ear positions and pupil dilation17. Others can track soil water content, solar irradiance, and weather conditions, then be used to predict the nitrogen response rate (NRR) of pasture dry matter and monitor soil conditions18.
RL agents could use these DTs to generate synthetic data for training, then find policies that recommend irrigation, lighting and nutrient dissemination to minimize resource-use whilst maximizing crop yield19.
Moreover, DTs may promote rewilding, sediment trapping and additional nature-based solutions for land management and restoration20, through rapid experimentation in ‘virtual farms’. In silico ‘what-if’ simulations could elicit further benefits, such as testing and identifying pathways to increase carbon sequestration in croplands and pastures, or using agro-forestry techniques, such as integrated green belts for wildfire prevention.
As in other domains, including water and electricity infrastructure, DTs can support predictive maintenance21, for instance, of irrigation systems in plantations to minimize food losses. In intensive controlled environment agriculture (CEA), such as commercial aeroponic greenhouses and hydroponic systems, DTs may be used in structure design and operations to suggest optimum light intensity, humidity, temperatures, CO2 concentrations and water-nutrient recycling.
Storage and transportation
Commodity chains that connect local produce to markets typically involve transit in freight trains and bulk carriers as well as temporary storage in terminal elevators. In rail, road and sea vessels, and in storage silos, cargos of grain are susceptible to mold, mustiness and early germination.
Ventilation management is essential to prevent dampness and fungal infestation, such as Aspergillus and penicillium that frequently deteriorate the quality of cereal bulks22. DTs already employed for improved HVAC systems design23 could be re-purposed to this end. In addition, real-time replicas of stationary elevators and vessels on voyage could track ventilation periods and moisture content of cargo as well as provide warning of mycotoxin contamination that warrants fumigation.
DTs can monitor fruit quality during inter-continental shipping24. Combining live temperature measurements with mechanistic models, such DTs can predict parts of the fruit that will perish before delivery. Amalgamated insights from many of these DTs can provide insights into transportation conditions that reduce food quality, informing new delivery strategies that limit food waste.
In cold chains of perishable produce, where fruit, vegetable, dairy, meat and seafood products are pre-cooled and provisionally stored in refrigerated facilities, computer simulations may advise on energy efficiency measures to reduce carbon emissions. Synchronized DTs can monitor food temperatures, humidity, delivery schedules, respiratory behavior, and grid carbon intensity; RL agents can then optimize the control of cooling equipment to draw power from the grid when carbon intensity is lowest to minimize emissions whilst maintaining food quality.
Paired with sensing technologies, DTs can be integrated across food processing and packaging facilities that convert agricultural commodities, such as corn or cattle, to ingredients and end-user food products, including tinned vegetables, meat cuts, ready meals and confectionery25.
Food loss and waste in this echelon are prevalent in both developed and developing regions, with implications for food security and the environment. In the UK, for example, food waste in this echelon stands at five megatonnes each year26.
Here, DTs can support industrial ecology approaches to prevent food loss, in the same way they have been used to enhance circular economy applications in construction manufacturing27. DTs can be deployed in smart manufacturing plants to monitor ingredient delivery schedules, plant throughput, ingredient wasteage, operator work schedules and demand forecasts. RL models can then be trained to manipulate manufacturing equipment to match food processing to expected demand whilst minimizing waste28.
Distribution and retail
Food distribution networks are significant contributors to global GHG emissions, with food retail alone responsible for ~0.3 gigatonnes of CO2 annually29. Food discarded in this echelon is considerable too, for example, with estimates suggesting 366 kilotonnes of food waste per year in the UK26. These losses are attributed to inefficient warehousing, hypermarkets and supermarkets operations including shelf management and failure to monitor and measure food waste30.
DTs that track construction-site logistics31 could be repurposed to mimic food distribution systems, and used to optimized delivery schedules minimizing carbon emissions and food wasteage. Such DTs could monitor the location of delivery vehicles across the road network, food inventory in retail stores, food embodied emissions traffic, weather and shelf-life of food in transit.
DTs can model the cold chain end-to-end to provide retailers with a better understanding of food quality when it arrives in-store32. Here, live temperature readings inform physics-based food models to track quality throughout distribution. By performing sensitivity analyzes on such models, and inferring optimized shipping conditions fruit shelf life can lengthen.
Given this state representation, RL agents used to optimize supply chain distribution to maximise producer profit could be repurposed to maximise resource efficiency33. Agents could synthesise policies that minimize food wasteage, and thus system-level emissions, by sending food to a retailer further from the distribution center with low inventory levels, rather than a closer store more likely to incur wasteage. Recent reviews suggest these simulations could further predict delays in supply chains, signs of food spoilage and potential food losses as well as recommend preventative measures34.
Where discard of food surplus is expected, the expansion of DTs to encompass networks of food re-distribution, such as community soup kitchens, can aid in waste mitigation and improving the nutritional security of vulnerable populations. Such expansion may also include growers to more effectively apportion and dispense unharvested produce.
Malnutrition, which currently afflicts over two billion people, arises from deficient, excessive or imbalanced consumption of macro- and micro-nutrients. Insufficient intake of iodine and iron, for instance, may lead to anaemia. Overconsumption of carbohydrates, for example, can result in increased risk of cardiovascular diseases.
One recent and emerging approach to the predicament of malnutrition is nutrigenetics. This field of research proposes that individuals’ genetic profile and microbiome determines their metabolism, nutrient requirements, predisposition to nutrition-related diseases such as type 2 diabetes, and response to dietary interventions35. To the extent that DTs could, in the future, simulate individual persons36by combining omics data, including nutrigenomics and metabolomics, and drawing on medical and lifestyle records, including via IoT wearable devices, virtual representations of humans could generate scenarios on the health effects of their food choices, customize dietary interventions and transform preventive healthcare thereby reducing malnutrition