Puff characterization using machine learning for an aerosol-generating device or system

Abstract

An aerosol-generating device includes a power supply for supplying power to generate the aerosol One or more sensors monitor one or more parameters associated with current operating conditions of the aerosol-generating device during operation of the aerosol-generating device. A storage device stores a machine learning model trained using a plurality of known target variable values over a range of operating conditions associated with the one or more parameters of the aerosol-generating device. A controller is coupled to the power supply, the one or more sensors, and the storage device. The controller uses the machine learning model to determine a current target variable based on the one or more parameters.

Description

[0001]- 1 - PUFF CHARACTERIZATION USING MACHINE LEARNING FOR AN AEROSOL- GENERATING DEVICE OR SYSTEM The present disclosure relates to a method of accurately determining a current target variable for an aerosol-generating device, and an aerosol-generating device utilizing a machine learning model to determine a current target variable to control the aerosol-generating device. In particular, the disclosure relates to an improved method utilizing a trained machine learning model to accurately determine puff volume as the current target variable for an aerosol-generating device. Aerosol-generating devices are typically configured to release an aerosol by heating a solid, a liquid, or an amorphous form of an aerosol-forming article. An exemplary aerosol- generating device includes an elongated main body with a power unit, a control unit, and a heating chamber to which an aerosol-forming article can be removably introduced to. The heating chamber typically includes a heating arrangement for heating the heating chamber. Aerosol-generating devices that generate an aerosol from an aerosol-forming substrate of an aerosol-forming article without requiring combustion of the aerosol-forming substrate are known. Such articles are often designated as “heat-not-burn” aerosol-forming articles, since an aerosol-forming substrate is heated to a relatively low temperature to induce the formation of an aerosol but prevent the combustion of material contained within the aerosol-forming substrate. These aerosol-forming articles having such substrates are combined with an aerosol-generating device to form an aerosol-generating system. Other known aerosol-generating devices are configured to vaporize a liquid having one or more active ingredients for inhalation, and such articles are often designated as “e-Vapor” or “vaping” devices. Some aerosol-generating devices are configured to provide usage sessions that have a finite duration. The duration of a usage session may be limited, for example, to approximate the experience of consuming a traditional cigarette. Some aerosol-generating devices are configured to be used with separate, consumable, aerosol-forming articles. Such aerosol-forming articles comprise an aerosol-forming substrate or substrates that are capable of releasing volatile compounds that can form an aerosol. Aerosol-forming substrates are commonly heated to form an aerosol. As the volatile compounds in an aerosol-forming substrates are depleted, the quality of the aerosol produced may deteriorate. Thus, some aerosol-generating devices are configured to limit the duration of the usage session to help prevent generation of a lower quality aerosol from a substantially depleted aerosol-forming substrate. - 2 - In some aerosol-generating devices, the duration of a usage session may be determined purely by time duration. One problem associated with setting a limit on a usage session purely based on time is that no account is taken of use behavior of a user. Thus, a user that takes a large number of puffs may deplete available aerosol-forming substrate from an aerosol-forming article within the duration of a usage session. In some aerosol-generating devices, the number of puffs taken by a user during a usage session is recorded and the duration of a usage session may be determined partially or completely based on the number of puffs taken by a user. As an example, an aerosol-generating device may be configured to produce aerosol from an aerosol- forming substrate during a usage session and the usage session may be terminated after a user has taken 14 puffs from the aerosol-forming substrate. Users taking 14 long puffs may still deplete available aerosol from the aerosol-forming substrate within their usage session, while users taking 14 short puffs may find their usage session is terminated before the available aerosol from the aerosol-forming substrate has been fully consumed. One way to determine the amount of aerosol that has been consumed from the substrate by a puff that is inhaled by the user is to use a value indicative of the electrical heating energy that is required to produce the aerosol for said puff. However, it has proven challenging to guarantee an acceptable level of error in the puff energy calculation using a method that allows puffs until the total volume budget is reached, given the strong variability of stick and environmental conditions in which the usage session could take place, and the limited availability of different types of sensor data that could help for such calculation. This variability and limited data have a strong impact on the reliability of the computation. Therefore, it is desirable to have an aerosol-generating device with improved puff volume estimation or puff detection to increase reliability resulting in decreasing the risk of wrong estimations and false positive puff detections. It would also be desirable to have an aerosol- generating device with increased reliability of the determination of the end of the usage session. It would also be desirable to have an aerosol-generating device with a finer resolution of puff volume estimation. It would also be desirable to have an aerosol-generating device that provides increased user satisfaction. The present invention relates to an aerosol-generating device configured to characterize a target variable using (or operationalizing) a machine learning model and one or more sensors disposed on the aerosol-generating device. In particular, the machine learning model was trained using known target variable values over a range of operating conditions and thereafter operationally transferred to the aerosol-generating device. - 3 - The present invention provides a method for producing a machine learning model for an aerosol-generating device for generating an aerosol from an aerosol-forming article. The method includes training the machine learning model using a plurality of known target variable values over a range of operating conditions associated with one or more parameters of the aerosol- generating device and transferring the machine learning model to a controller of the aerosol- generating device. The machine learning model is configured to receive as input one or more measured parameter values of the aerosol-generating device associated with current operating conditions and output a current target variable. The plurality of different properties and operating conditions include different aerosol- forming articles, different aerosol-forming substrates, different operating power, different environmental humidity levels, different environmental pressure levels, different environmental temperature levels, different internal electronics temperature levels of the aerosol-generating device, and different aerosol-forming article states. The plurality of known target variable values are known puff volumes and the current target variable is a current puff volume of a puff that has been generated by the aerosol-generating device. The one or more parameters associated with the aerosol device comprise one or more of a voltage, a current, a power, and a temperature associated with the aerosol-generating device. The one or more parameters may be determined using one or more sensors disposed in or on the aerosol-generating device. For example, temperatures may be measured, or data indicative of temperatures may be captured using temperature sensors disposed in or proximate to the heating cavity of the device, for example in the airflow path or a side wall of the airflow path upstream of the aerosol-forming substrate, to thereby capture data indicative of the temperature of elements of the airflow path. In some examples, data indicative of a temperature of the heating process during the usage session is captured by a temperature sensor. In some examples, data indicative of a temperature of the control electronics, for example the temperature of the electronic circuits or circuit board is captured, to have data that is indicative of the operational status of the aerosol-generating device. For example, a heating status of the usage session, or a temperature of the aerosol-forming substrate is determined, as the operation of the heater or substrate may impact a temperature inside the aerosol-generating device. For example, as many electronic circuits, such as microcontrollers or other circuitry already have internal temperature sensing, this data can be used for such purpose without the need and costs of adding an additional temperature sensor. Other types of parameters may include a current supply voltage supplied to the heater, a current consumed by the heater, current from the battery, and a heater power, for example. Any - 4 - of these parameters may be used individually or in combination with one more other parameters to train the machine learning model. According to an aspect of the present invention, an aerosol-generating device is provided for generating an aerosol from an aerosol-forming article. The aerosol-generating device includes a power supply for supplying power to generate the aerosol. Preferably, one or more sensors monitor one or more parameters, respectively, associated with current operating conditions of the aerosol-generating device during operation of the aerosol-generating device. A storage device stores a machine learning model. A controller is coupled to the power supply, the one or more sensors, and the storage device. The controller uses the machine learning model to determine a current target variable based on the one or more parameters. According to an aspect of the present invention, an aerosol-generating device generates an aerosol from an aerosol-forming article having a solid aerosol-forming substrate. The aerosol- generating device includes a heater for heating the solid aerosol-forming substrate of the aerosol- forming article during a usage session. The usage session having a duration allowing for taking several puffs. A power supply supplies power to the heater to generate the aerosol. One or more sensors are included for monitoring one or more parameters associated with current operating conditions of the aerosol-generating device during the usage session of the aerosol-generating device. A storage device stores a machine learning model trained using a plurality of known target variable values over a range of operating conditions associated with the one or more parameters of the aerosol-generating device. A controller is coupled to the power supply, the one or more sensors, and the storage device. The controller uses the machine learning model to determine a current target variable related to a puff based on the one or more parameters. The one or more sensors comprise one or more of a power sensor, for example but not limited to a current sensor, a voltage sensor, or both, one or more temperature sensors, and optionally an accelerometer. The one or more sensors may exclude a flow sensor. The one or more sensors may exclude a temperature sensor that measures the outside or environmental temperature. The one or more of sensors may exclude expensive sensors or sensors that may not be able to operate reliably for long-term use. For example, the one or more sensors may exclude a humidity sensor that measures the outside or environmental humidity, as it may be possible that, depending on its arrangement, it is easily obstructed and not able to measure the humidity reliably. The one or more sensors may exclude a flow sensor that could directly gather data on the volumetric flow of the air in the airpath, as it could be clogged by particles or substances in the airflow during long term use, depending on the sensing technology used. The current target variable can control a control operation of the aerosol-generating device. The - 5 - aerosol-generating device is configured to generate aerosol, for example with one or more puffs, during a usage session. The controller is configured to determine a start of the usage session, monitor the one or more parameters associated with current operating conditions of the aerosol- generating device during the usage session, and use the current target variable to determine an end of the usage session. The plurality of known target variable values may be known puff volumes, and the current target variable is a current puff volume of the aerosol-generating device. Using the current target variable to determine the end of the usage session includes determining a cumulative puff volume measured from the start of the usage session and use the cumulative puff volume to determine the end of the usage session. The controller ends the usage session, and thereby stops or winds down the heating of the aerosol-forming substrate, when the cumulative puff volume reaches a predefined threshold. An aerosol-generating device using a machine learning model that is operationalized to characterize target variables advantageously provides improved target variable estimation (for example., puff volume and also puff detection (timing, for example, a start time and an end time of a puff taken by a user)). The machine learning model provides more accurate estimations in various operating conditions without the need for having complicated airflow measurement sensors that can clog or otherwise be rendered inoperative due to particle accumulation and associated unreliable data processing for improving detection and estimation accuracy. Also, the need for other sensors may be obviated, that may add costs to the aerosol-generating device or may provide inaccurate data to operate in the changing real-world environment and may be difficult to implement. The trained machine learning model that is operationalized allows for a finer resolution in the determination of the target variable. A system or device that does not use a machine learning approach as described herein may have difficulty adapting to different stick conditions (dry or humid, material or composition variations from manufacturing, for example) because, for example, more electric power may be needed to heat up a stick with a higher water content than a dry stick. In the aerosol-generating device described herein energy measured during a puff may not only vary depending on the actual puff volume but also based on the substrate conditions of the aerosol-forming substrate, for example, humid or dry conditions, age of the substrate, which changes the power absorption. An aerosol-generating device using or operationalizing a trained machine learning model to characterize puff volume and puff timing advantageously provides increased reliability of the determination of a usage session end time. The increased reliability can prevent a usage session - 6 - from ending prematurely where aerosol-forming material is still available, or from extending beyond a time after the substrate has been fully consumed. Advantageously, according to an aspect of the current invention, an improved target variable estimation method and device is provided while preventing higher costs by using a machine learning model built offline that does not change over time, and by using standard parameters or values that do not require additional specific sensors. According to an aspect of the invention, the method and device allows for the use of a machine learning model that is created offline and uploaded to the aerosol-generating device upon manufacture or assembly, for example. The use of an unchanging machine learning model on the aerosol-generating device to characterize the target variables allows for an accurate estimation of the target variables without requiring sensors that are not already disposed in the aerosol-generating device. This prevents a higher costs and higher operational complexity that are associated with an aerosol-generating device having additional added sensors. The inclusion of a flow sensor to measure the puff volume may be unreasonable as it would increase costs, encumbrance, and complexity of the device (let alone the fact that a varying air infusion in the consumable would affect the amount of air going through such puff sensor). It will be appreciated that a direct puff volume measurement using a flow sensor, for example, would be cumbersome in terms of device complexity and adding an additional element that could be subject to operational failure. A flow sensor may clog with slurry, particles, or debris (if putting the device into a pocket full of dust). In terms of design, use of a flow sensor is particularly difficult because the flow sensor requires a non-negligible space. According to the present solution, there is no need of any additional hardware. The solution provided herein provides a layer of software on top of the existing heating control algorithm, which is already capable of measuring various parameters such as power, voltage, current, and temperature inside or outside the heating cavity using sensors already present on the device. In an example, a temperature sensor can include a thermistor that can detect a value indicative of a temperature change that is caused by the airflow caused by the user taking a puff. The airflow causes a cooling effect to the walls of the heating cavity, and this temperature change can be sensed by the thermistor. According to various examples, the temperature sensor is located outside the heating cavity prevents the sensor from being damaged. At the same time, this arrangement unexpectedly - 7 - still allows for detecting a user's puff fast enough to enable accurate temperature control, notwithstanding a possible time delay between the occurrence of a temperature drop of the air flow in the cavity and its detectability at the position of the temperature sensor outside the cavity. In particular, the temperature sensor may be configured to detect a temperature change of air flow in the heating cavity via thermal conduction through a wall member of the device which forms a least a portion of the heating cavity. The temperature sensor preferably detects a change of the temperature of the cavity at the test spot which is caused by a change of the temperature of the air flow passing along the inner surface of the cavity at the place of the test spot. Preferably, the temperature sensor may comprise at least one thermistor. As compared to other types of temperature sensors, for example resistance temperature sensors, thermistors advantageously achieve a greater precision within a limited temperature range. Thermistors also provide a proper temperature response because the resistance of thermistors strongly depends on temperature more so than in standard resistors. In addition, thermistors are well suited for point sensing as being capable of achieving high accuracy in measuring the temperature at a specific point. This is particularly advantageous for measuring the temperature of air flow close to the test spot on the inner surface of the receiving cavity. Nor example, the thermistor may be a negative temperature coefficient (NTC) thermistor. A NTC thermistor comprises a resistance which decreases as temperature rises. NTC thermistors are particularly suitable for monitoring small changes in temperature as those occurring during a user's puff. This is due to the resistance of the material of a NTC thermistor being linearly proportional to the temperature over small changes in temperature. A method, device and system of puff detection using a temperature sensor located outside of the heating cavity is described in more detail in U.S. Pat. Pub. No. 2022 / 0211113, which is incorporated by reference in its entirety. According to some examples, a value indicative of the temperature change that is caused by the airflow during a puff is measured and this value is used by the machine learning model to determine puff volume or other target variable. The temperature change may be determined at various locations within or on the aerosol-generating device. For example, the temperature change may be detected proximate or on a wall, side wall, or element of the airflow path, a wall, side wall, or element of the heating cavity, or both a wall, side wall, or element of the airflow path and the heating cavity. Although the exact same effect may the drop of the temperature of the heater (the temperature-based control of the heater temperature leads to power peaks to compensate for the cooling effect of the airflow during a puff), the cooling effect that is caused by the airflow to walls of the airflow path or the walls of the heating or heating cavity walls may happen first. According - 8 - to another aspect of the present invention, a method of operating an aerosol-generating device is provided, described herein. The method includes monitoring one or more parameters associated with current operating conditions of the aerosol-generating device during operation of the aerosol- generating device and using a machine learning model to determine a current target variable based on the one or more parameters. The machine learning model trained using a plurality of known target variable values over a range of operating conditions associated with the one or more parameters of the aerosol-generating device. Advantageously, according to an aspect of the present invention, the method allows for increased user satisfaction. Accurately determining the duration of a usage session enhances user satisfaction as the usage session is not terminated prematurely or extend beyond a time that the substrate has been fully consumed. As used herein, the singular forms “a,” “an,” and “the” also encompass embodiments having plural referents, unless the content clearly dictates otherwise. As used herein, “have”, “having”, “include”, “including”, “comprise”, “comprising” or the like are used in their open-ended sense, and generally mean “including, but not limited to”. It will be understood that “consisting essentially of”, “consisting of”, and the like are subsumed in “comprising,” and the like. The words “preferred” and “preferably” refer to embodiments of the invention that may afford certain benefits, under certain circumstances. However, other embodiments may also be preferred, under the same or other circumstances. Furthermore, the recitation of one or more preferred embodiments does not imply that other embodiments are not useful and is not intended to exclude other embodiments from the scope of the disclosure, including the claims. Any direction referred to herein such as “top”, “bottom”, “left”, “right”, upper”, “lower”, and other directions or orientations are described herein for clarity and brevity but are not intended to be limiting of an actual device or system. Devices and systems described herein may be used in a number of directions and orientations. As used herein, “downstream” and “proximal” mean the mouthpiece end of the aerosol- generating device. “Downstream” and “proximal” mean the end of the aerosol-generating device intended to be contacted by the mouth of a user. “Upstream” and “distal” mean the opposite end of the aerosol-generating device. - 9 - As used herein, “tobacco” means plant material, such as leaves, stems, or other portions of any of several plants belonging to the genus Nicotiana, such as of the species N. tabacum. Preferably, tobacco includes leaves, stems, or leaves and stems. As used herein, a “controller” is one or more hardware devices, one or more software or firmware programs, or one or more hardware devices and software or firmware programs that manages or directs flow of data between two or more entities. The controller may include a memory, an Application-Specific Integrated Circuit (ASIC) state machine, a digital signal processor, a gate array, a microprocessor, or equivalent discrete or integrated logic circuitry. A controller may include memory that contains instructions that cause one or more components of the circuitry to carry out a function of the controller. Functions attributable to a controller in this disclosure may be embodied as one or more of software, firmware, and hardware. The controller may include a microprocessor. The operation of one or more controller of a system may be coordinated by an overarching system controller. The term “aerosol” is used here to refer to a suspension of solid particles or liquid droplets, or a combination of solid particles and liquid droplets in a gas. The gas may be air. The solid particles or liquid droplets may comprise one or more volatile flavor compounds. Aerosol may be visible or invisible. Aerosol may include substances that are ordinarily liquid or solid at room temperature. Aerosol may include substances that are ordinarily liquid or solid at room temperature, in combination with solid particles or in combination with liquid droplets or in combination with both solid particles and liquid droplets. The aerosol preferably comprises nicotine. The term “aerosol-generating device” is used here to refer to any device configured to be used or utilized with an aerosol-forming article that releases volatile compounds to form an aerosol that may be inhaled by a user. The aerosol-generating device may be interfaced with the aerosol-forming article comprising the aerosol-forming article. The term “aerosol-forming article” is used herein to refer to a disposable product capable of including (for example, holding, containing, having, or storing) aerosol-forming substrate. An aerosol-forming article may be capable of removably interfacing, or docking, or mating with an aerosol-generating device. This allows the aerosol-generating device to generate aerosol from the aerosol-forming substrate of the aerosol-forming article. A method for producing a machine learning model for an aerosol-generating device for generating an aerosol from an aerosol-forming article includes training the machine learning - 10 - model using a plurality of known target variable values over a range of operating conditions associated with one or more parameters of the aerosol-generating device and transferring the machine learning model to a controller of the aerosol-generating device for operationalization and data inference. The machine learning model receives as input one or more measured parameter values from one or more sensors of the aerosol-generating device associated with current operating conditions and output a current target variable. The plurality of different properties includes different aerosol-forming articles, different aerosol-forming substrates, different operating power, different humidity levels, different environmental pressure levels, different temperature levels, and different aerosol-forming article states. The plurality of known target variable values is known puff volumes, and the current target variable is a current puff volume generated by the aerosol-generating device, or a puff detection (timing) generated by the aerosol-generating device. It is also possible that other current target variables that can be estimated or detected by the machine learning model, for example but not limited to; an overall depletion level of the aerosol-forming article during the usage session, for example based on estimated cumulative puff volume during a usage session, a user identification based on puffing habits that can be expressed by puff volume, puff volume evolution during a single or multiple puffs, puff frequency and puff timing during a usage session, motion departed to the aerosol-generating device during a usage session and or one or more puffs, an indication or estimation of an aerosol-temperature for example a temperature of the aerosol exiting the aerosol-forming article, for example to determine whether the aerosol will be too hot for inhalation, to detect or at least indicate presence or absence of the hot aerosol effect. In one example, the plurality of known target variable values are known puff volumes and the current target variable is at least one of a current puff volume of the aerosol-generating device, or a temperature of the aerosol exiting the aerosol-forming article for consumption. The one or more parameters associated with the aerosol-generating device can comprise one or more of a voltage, a current, a power, and one or more temperatures associated with the aerosol-generating device, acceleration and orientation data from an accelerometer or inertial measurement unit (IMU). An aerosol-generating device for generating an aerosol from an aerosol-forming substrate of the aerosol-forming article includes a power supply for supplying power to a heater to generate the aerosol. One or more sensors or a plurality of sensors monitor one or more parameters associated with current operating conditions of the aerosol-generating device during operation of the heater of the aerosol-generating device. A storage device is configured to store a machine learning model. A controller is coupled to the power supply, the one or more sensors, and the storage device. - 11 - The controller uses the operationalized and trained machine learning model to determine a current target variable based on the one or more parameters. The one more parameters may be representative of power supplied by the power supply. For example, a power supply may supply power to maintain a heater at a predetermined temperature during a usage session. If a user puffs on the device to generate an aerosol, the heater cools and a greater amount of power is required to maintain the heater at the predetermined temperature. Thus, by monitoring a parameter representative of power supplied by the power supply, a value indicative of real time aerosol generation may be recorded.The one or more sensors includes one or more of a power sensor, for example measuring a value indicative of the current consumed by the heater, a voltage supplied to the heater, or an electric power supplied to the heater, one or more temperature sensors, and an accelerometer. In some examples, a hot aerosol effect can be detected during a usage session, or even during a puff itself, using the trained machine learning model that has been operationalized on the aerosol-generating device. For example, a prediction of a hot aerosol effect can be made, for example by estimating a probability before a puff that the inhaled aerosol will be or may be above a certain temperature threshold. This may be advantageous as the hot aerosol effect may be difficult to directly measure solely with a temperature sensor, as it is difficult to measure the temperature of the aerosol that exits the aerosol-forming article. This permits the controller of the aerosol-generating device to perform mitigating measures against the hot aerosol effect, as described in International Patent Publication No. WO2023 / 217937, this reference herewith incorporated by reference in its entirety. It is also possible to have a visual indicator or user interface on the aerosol-generating device that can be operated by the controller of the aerosol- generating device, to warn the user about a potentially occurring hot aerosol effect. In some examples, only one sensor of the one or more sensors is used to the puff volume detection method. For example, only an output from a temperature sensor that measures the temperature of the heating cavity may be used by the machine learning model to estimate the target variable. In some examples, only a current sensor is used. The current may be measured over a shunt resistor, for example. The unused sensors may still be present in the aerosol- generating device. In some examples, the unused sensors are not present on the aerosol- generating device. Preferably, the one or more sensors do not include any added sensors such as humidity sensors that could measure a humidity of the external environment, or other sensors that can clog or otherwise be rendered inoperative due to particle accumulation. In some examples, an aerosol-generating device may include additional sensors such as a humidity sensor or a flow sensor, but data from one or more of these additional sensors is not used in the estimation of the target variable. - 12 - The current target variable is configured to control operation of the aerosol-generating device. The aerosol-generating device generates aerosol during a usage session. The controller determines a start of the usage session, monitors the one or more parameters associated with current operating conditions of the aerosol-generating device during the usage session, for example but not limited to the one or more puffs that will be taken by the user, and uses the current target variable to determine an end of the usage session. The plurality of known target variable values may be known puff volumes, and the current target variable is a current puff volume of the aerosol-generating device. Using the current target variable to determine the end of the usage session includes determining a cumulative puff volume measured from the start of the usage session and use the cumulative puff volume to determine the end of the usage session. The controller can end the usage session when the cumulative puff volume reaches a predefined threshold. The devices and methods described herein allows for tailoring the length of a usage session based on the puffing style of the user. This is accomplished by allocating a budget aerosol volume for each usage session and calculating for each puff the volume of the aerosol generated: once the total volume allowed has been reached, the usage session comes to an end. This means that if a user has a very strong puffing style, he or she may be able to draw less puffs during a usage session versus a user making milder puffs. The computation of the aerosol volume may be at least partially based on a power signal analysis. For example, when a puff is drawn, a power peak compensates the temperature drop experienced by the heating element (due to the airflow cooling), and such power peak is analysed to calculate the associated aerosol volume. Once a cumulative volume of aerosol is reached, the usage session ends. In an example, a user inserts an aerosol-forming article into the heating chamber of the aerosol-generating device and can initiates a usage session by actuating an element of the user interface, for example the user button or by an automated start after the aerosol-generating device. This indicates the start of the usage session. Power is supplied from a battery to the heater of the aerosol-generating device until the heater provides for a predetermined operating temperature. The operating temperature may include one or more temperatures associated with the operation of the aerosol device. For example, the operating temperature may include one or more of a heater temperature, a heating chamber temperature, or an aerosol-forming substrate temperature. This temperature may be such that the aerosol-forming substrate can aerosolize one or more ingredients upon taking of a puff. - 13 - The heater may be active for the duration of the user session taking several puffs. The user session may include at least 5 puffs, or at least 10 puffs, or from 5 to 50 puffs, or from 5 to 25 puffs. The user session may have a time duration of at least 30 seconds, or at least 60 seconds, or at least 120 seconds, or from 30 seconds to 1000 seconds. The power signal of the power supplied to the heater can be monitored by the controller. The user then takes a puff. When the user puffs, the heater is cooled because of the airflow with air from the external environment flowing past the heater thereby imparting a cooling effect. Thus, the electric power that needs to be supplied to the heater to maintain the desired operating temperature increases. The power supplied increases and the correct temperature is maintained. The presence of a user puff may be detected by analysing the power signal. Various parameters may be measured to determine the power signal. For example, a current, a voltage, a power that is applied to the heater, or one or more temperatures associated with operation of the aerosol-generating device may be used to determine the power signal. For example, in a variant, a value that is indicative of the electric power that is used by the temperature-controlled heater is monitored, for example by current sensing, to have information that is indicative of the start time and end time of the puff of the user, and to have information that is indicate of the volume of the puff. For example, a method, system, and device to detect a puff based on the power signal is described in U.S. Pat. Pub. No.2015 / 0230521, this reference hereby incorporated by reference in its entirety. A puff start point, and a puff end point may be determined by means of this analysis. The energy of the detected puff is then calculated, and the volume of aerosol generated during the puff is also calculated and added to a cumulative total of volume generated during the usage session. If the cumulative total volume equals or exceeds the predetermined maximum permissible aerosol volume for the usage session (for example but not limited to an aerosol volume of 660 ml) the usage session is ended, for example by stopping or ramping down the heater temperature. If the cumulative total volume does not equal or exceeds the predetermined maximum permissible aerosol volume for the usage session, then the session can remain active, and the user may take another puff. The usage session remains active until the user has generated the maximum permissible aerosol volume or until a maximum time threshold is reached. The improved accuracy of the estimation of puff volume is accomplished by using a machine learning model. For example, the accuracy can be improved in particular when facing variable environment conditions and different levels of air infusion. As a non-limiting example of - 14 - the air infusion, the relative amount of air going directly in the condensation chamber of the consumable rather than passing through the tobacco plug containing the susceptor, can vary between different aerosol-forming articles. Also, a part of the airpath may be clogged or otherwise impacted and therefore provide for variable air infusion between different devices and usage sessions. The machine learning model establishes correlations between different parameters available in device data (training sets) and actual puff volume, to take into account these variations. According to yet another aspect of the present invention, a method of training a machine learning model is provided. During the learning phase that is applied to the model, the machine learning model is fed with data collected from a high number of usage sessions, for example performed with a smoking or inhalation machine in a controlled manner, labelled for the aerosol volume they represent. For example, a usage session with twelve (12) puffs each one of 35 ml (as an example) repeated in a plurality of different conditions such as different air infusion sticks, different power states of the device, and different substrate states (different agglomerations of cast leaf, different humidity levels, for example), different environmental temperatures, different environmental humidity levels, different water content of the aerosol-forming substrates of the aerosol-forming articles. In some examples, the method includes a learning step where human puffs by a specific user or a plurality of users are taken during a high number of usage sessions, with the goal to learn or train the machine learning model. This can be done to complement the learning step done by the inhalation machine or done as a step that replaces the data generated by the puff taken from the inhalation machine. The puffs taken by human inhalation can be measured and characterized accurately with the use of a reference flow meter that is operatively connected to a data processing device and provides for flow measurement data. The reference flow meter may be fluidically arranged between a mouth of the user and the aerosol-generating device. More specifically, the flow meter may be disposed on the mouthpiece side of the aerosol-forming article of the aerosol-generating device. During this learning step, it is possible to gather precise data that characterizes the human puff taken during usage session. With the reference flow meter, it is possible to provide for a highly accurate reference value for the puff volume for each puff, that thereafter can be used to label each user puffs before using these values to train or learn the machine learning model. This adds to the data obtained using calibrated puffs (for example, known volume) with automated puff machines. Puffs for such data collection are executed in a controlled manner (with smoking machines puffing exactly at a certain known volume, or by human inhalation with a flow meter, or both). The - 15 - training data provides usage sessions with variable conditions, but with puffs that are characterized by a known volume, for example 35 ml. Then, the same data set is again collected for a different bucket for a different puff volume, for example 55 ml. Again, 55 ml puffs are repeated in a plurality of different conditions, for example, sticks with different air infusion, different power devices (high power or low power depending on the health status of the battery or its charge level), dry sticks, humid sticks, such that most or all possible spectrum of conditions for such puff volume bucket are covered. The same type of data collection is repeated for many buckets (35ml, 45ml, 55ml, 65ml and so on). The higher the number of buckets, the higher will be the resulting firmware accuracy. For each puff at a certain volume bucket all available characterizing data is fed into the machine learning model such as voltage, current, power, temperature of the puff sensor, a temperature of the cavity of the aerosol-generating device, and a temperature of the printed circuit board (PCB) of the aerosol-generating device. Each puff is isolated and characterizing data is sampled during the time of the puff and stored. The collected data may be restricted to a certain time period before or after a puff. For example, the collected data is restricted from 1, 2, 3, 4, or 5 seconds before the puff and up to 1, 2, 3, 4, 5, or 6 seconds after its end so that the most significant data is used. Data collected in other scenarios may actually distract the machine learning and potentially lead to a loss of accuracy. In this way, the machine learning model is able to create correlations between elements that are present in all puffs at a certain volume, no matter the air infusion, no matter the substrate conditions such that when the user delivers (for example) a 35 ml puff, the system is capable of quantifying the puff at a given volume (for example 35 ml). This is possible because the trained firmware found, during the machine learning off-line phase, a set of features of the data which is typical for the 35 ml puffs. The machine learning model and the method of training the machine learning model, according to examples described herein, may not be intended to be learning in the field, that is during the usage sessions or sessions performed by the user, but created in the lab or other controlled environment based on a strictly controlled training set and uploaded to the aerosol- generating device upon manufacture, for example. The training set may use data collected from sensors that are not available at the aerosol-generating device that is used by the user to further refine the data set, allowing to avoid local minima and other convergence issues of the data set. When building the machine learning model, characteristics of all puffs in different environmental conditions are extracted and a data repository or data base is created. The collected data may be fed to one or more different machine learning algorithms to create the trained firmware. For example, the machine learning algorithms may include one or more of a neural network for - 16 - example but not limited to an Artificial Neural Network (ANN), Convolutional Neural Network (CNN), Convolutional Recurrent Neural Network (CRNN) and Recurrent Neural Network (RNN), k-nearest neighbors algorithm (k-NN), support vector machines (SVM), a linear regression model, a decision tree, a decision forest, a random forest (RF), multi-layer perceptrons (MLP) fuzzy logic, correlation model, a classification model, and a clustering model. Once the machine learning model is created using the training data, it can be loaded into the firmware of the aerosol- generating device that is commercially available to the user. The result is a tool or software module, which ultimately can be considered as a black box sub-firmware configured to characterize the puff, which is to associate an existence of a puff or a volume to each puff drawn by the user. Building a machine learning model may involve significant computational resources running a considerable amount of data. The output or result of the method of training a machine learning model can include a firmware or other software code, which receives as input puff-related data from sensors already present on the aerosol-generating device (voltage, current, power, temperature of the heating chamber, temperature of the heater, and the like) and delivers, for each puff, the information relative to its volume. During operation of the aerosol-generating device, the device firmware identifies the occurrence of a puff. Then, it collects data associated to the puff that is indicative of power consumption by the heater and provides such package to the tool for determining the puff volume. The machine learning model (that is, the trained firmware) on the aerosol-generating device can be operationalized or put into production and is capable, during the usage session, to receive as input data samples during the occurrence of a puff (such as voltage, current, power, and temperature, for example) and return as output an estimation of the associated puff volume. In this sense, the trained machine learning model is used to generate predictions on a current target variable, for example but not limited to puff volume, on unseen real-world data from the one or more sensors of the aerosol-generating device in the operational environment, thereby performing model inference. Generally, inference in machine learning is the process where the trained machine model is used to make predictions or classifications on new, unseen data. In this context, the unseen data could be a test data set, or data coming from the real-world one or more sensors of the aerosol-generating device. Data for each puff includes all possible variables which can be measured during the puff, and that is the voltage, the current, the power, temperature of the thermistor (used as puff sensor), and the like. An aspect of such data collection is that no matter how much the scenarios can vary - 17 - in terms of stick conditions and the like, each plurality of data sets is associated to a precise puff volume bucket. Therefore, the machine learning model is trained based on the fact that data varies (because of the difference scenarios used) but they all return puffs characterized by the same aerosol volume. The aerosol volume in the training data is controlled because of a defined process or method gathering the training data, for example where a series of puffs are carried out by an automated device or system in a controlled manner, or a series of puffs are taken by a user during the training with the use of a flow meter for puff labelling. Accordingly, the machine learning algorithm is capable of establishing correlations between the data which are used during real device operations to make puff volume predictions. Once the data is collected, it is then fed to the machine learning algorithm to be trained and to ultimately return a “trained” software module that can be uploaded as a firmware capable of, during operation of the device, receive as input values samples during the puff (such as current, voltage and the like) using sensors already present on the aerosol-generating device and output a prediction of the aerosol volume of the puff. It is to be understood that while characterizing a puff volume is used as an example herein of a target variable, other target variable associated with an operation of the aerosol-generating device may be used. For example, the target variable may include a puff start time, puff end time, puff duration, airflow velocity during a puff, for example an air velocity profile over time, air flow volume during a puff, for example an air flow volume profile over time. In some examples, multiple target variables are determined during use of the aerosol-generating device based on the machine learning model Below there is provided a non-exhaustive list of non-limiting examples. Any one or more of the features of these examples may be combined with any one or more features of another example or embodiment described herein. Example Ex1: A method for producing a machine learning model for an aerosol-generating device for generating an aerosol from an aerosol-forming article, the method comprising: training the machine learning model using a plurality of known target variable values over a range of operating conditions associated with one or more parameters of the aerosol-generating device; and transferring the machine learning model to a controller of the aerosol-generating device, the machine learning model configured to receive as input one or more measured parameter values of the aerosol-generating device associated with current operating conditions and output a current target variable. - 18 - Example Ex2: The method of Ex1, wherein training the machine learning model comprises training the machine learning model using data collected from aerosol-generating devices with a plurality of different properties and a plurality of different operating conditions. Example Ex3: The method of Ex 2, wherein the plurality of different properties and a plurality of different operating conditions comprise different aerosol-forming articles, different operating power, different humidity levels, different environmental pressure levels, different temperature levels, and different aerosol-forming article states. Example Ex4: The method of any preceding example, wherein the machine learning model comprises one or more of an Artificial Neural Network (ANN), Convolutional Neural Network (CNN), Convolutional Recurrent Neural Network (CRNN), Recurrent Neural Network (RNN), k- nearest neighbors algorithm (k-NN), support vector machines (SVM), a linear regression model, a decision tree, a decision forest, a random forest (RF), multi-layer perceptrons (MLP), fuzzy logic, correlation model, a classification model, and a clustering model. Example Ex5: The method of any preceding example, wherein the plurality of known target variable values are known puff volumes and the current target variable is at least one of a current puff volume of the aerosol-generating device, a temperature of the aerosol exiting the aerosol- forming article for consumption. Example Ex6: The method any preceding example wherein the plurality of known target variable values are known puff start times and the current target variable is a current puff start time of the aerosol-generating device. Example Ex7: The method any preceding example wherein the plurality of known target variable values are known puff end times and the current target variable is a current puff end time of the aerosol-generating device. Example Ex8: The method of any preceding example, wherein the one or more parameters associated with the aerosol device comprise one or more of a voltage, a current, a power, and a temperature of an airflow channel of the aerosol-generating device. Example Ex9: The method of Ex8, wherein the one or more temperatures associated with the aerosol-generating device comprise one or both of a cavity temperature of the aerosol-generating device and a temperature of a printed circuit board (PCB) of the aerosol-generating device. - 19 - Example Ex10: An aerosol-generating device for generating an aerosol from an aerosol-forming article, the aerosol-generating device comprising: a power supply for supplying power to generate the aerosol; one or more sensors for monitoring one or more parameters associated with current operating conditions of the aerosol-generating device during operation of the aerosol-generating device; a storage device for storing a machine learning model trained using a plurality of known target variable values over a range of operating conditions associated with the one or more parameters of the aerosol-generating device; and a controller coupled to the power supply, the one or more sensors, and the storage device, the controller for using the machine learning model to determine a current target variable based on the one or more parameters. Example Ex11: The aerosol-generating device of Ex10, wherein the one or more sensors comprises one or more of a power sensor, for example, but not limited to a current sensor or a voltage sensor one or more temperature sensors, and an accelerometer. Example Ex12: The aerosol-generating device of Ex11, wherein the one or more temperature sensors are configured to measure one or more temperatures associated with the aerosol- generating device, the one or more temperatures associated with the aerosol-generating device comprising one or both of a cavity temperature of the aerosol-generating device and a temperature of a printed circuit board (PCB) of the aerosol-generating device. Example Ex13: The aerosol-generating device of any of Ex10 through Ex12, wherein the aerosol- forming article comprises an aerosol-forming substrate and the aerosol-generating device comprises a heater to heat the aerosol-forming substrate. Example Ex14: The aerosol-generating device of Ex13, wherein the one or more sensors comprises one or both of a sensor that can provide for data that is indicative of a temperature caused by a cooling effect of an airflow to an element that forms an airflow path and a sensor that can provide for data that is indicative of a power consumption of the heater. Example Ex15. The aerosol-generating device of Ex13 or Ex14, wherein the sensor that provides data indicative of the temperature includes a temperature sensor that can measure a temperature of a side wall of the airflow path upstream of the aerosol-forming substrate. Example Ex16. The aerosol-generating device of any of Ex13 through Ex15, wherein the sensor that provides data indicative of the power consumption of the heater includes one or both of a voltage and a current sensor. - 20 - Example Ex17: The aerosol-generating device of any of Ex10 through Ex16, wherein the one or more sensors do not include a flow sensor. Example Ex18: The aerosol-generating device of any of Ex10 through Ex17, wherein the one or more sensors do not include a humidity sensor that measures environmental humidity. Example Ex19: The aerosol-generating device of any of Ex10 through Ex18, wherein the current target variable controls operation of the aerosol-generating device. Example Ex20: The aerosol-generating device of any of Ex10 through Ex19, wherein the aerosol- generating device generates aerosol during a usage session, wherein the controller: determines a start of the usage session; monitors the one or more parameters associated with current operating conditions of the aerosol-generating device during the usage session, and uses the current target variable to determine an end of the usage session, for example, stopping or ramping down the heating by the one or more heaters. Example Ex21: The aerosol-generating device of Ex20 wherein the plurality of known target variable values are known puff volumes and the current target variable is a current puff volume of the aerosol-generating device, and wherein using the current target variable to determine the end of the usage session comprises, determining a cumulative puff volume measured from the start of the usage session; and using the cumulative puff volume to determine the end of the usage session. Example Ex22: The aerosol-generating device of Ex21, wherein the controller ends the usage session when the cumulative puff volume reaches a predefined threshold. Example Ex23: A method of operating an aerosol-generating device, comprising: monitoring one or more parameters associated with current operating conditions of the aerosol-generating device during operation of the aerosol-generating device; using a machine learning model to determine a current target variable based on the one or more parameters, the machine learning model trained using a plurality of known target variable values over a range of operating conditions associated with the one or more parameters of the aerosol-generating device. Example Ex24: The method of Ex23, wherein the one or more sensors comprises one or more of a power sensor, one or more temperature sensors, and an accelerometer. - 21 - Example Ex25: The method of Ex24, wherein the one or more temperatures associated with the aerosol-generating device comprise one or both of a cavity temperature of the aerosol-generating device and a temperature of a printed circuit board (PCB) of the aerosol-generating device. Example Ex26: The method of any of Ex23 through Ex25, wherein the one or more sensors do not include one or both of a flow sensor and a humidity sensor. Example Ex27: The method of any of Ex23 through Ex26, wherein the current target variable controls operation of the aerosol-generating device. Example Ex28: The method of any of Ex23 through Ex27, wherein the aerosol-generating device generates aerosol during a usage session, wherein the controller: determines a start of the usage session; monitors the one or more parameters associated with current operating conditions of the aerosol-generating device during the usage session, uses the current target variable to determine an end of the usage session. Example Ex29: The method of Ex28 wherein the plurality of known target variable values are known puff volumes and the current target variable is a current puff volume of the aerosol- generating device, and wherein using the current target variable to determine the end of the usage session comprises, determining a cumulative puff volume measured from the start of the usage session; and using the cumulative puff volume to determine the end of the usage session. Example Ex30: The method of Ex29, wherein the controller ends the usage session when the cumulative puff volume reaches a predefined threshold. Example Ex31: The aerosol-generating device of any preceding claim, wherein the aerosol- forming article is a heat-not-burn aerosol-forming article. Example Ex32: The aerosol-generating device of any preceding claim, wherein the aerosol- generating device is a resistive heating aerosol-generating device. Example Ex33: The aerosol-generating device of any preceding claim, wherein the aerosol- generating device is an inductive heating aerosol-generating device. Example Ex34: The aerosol-generating device of any preceding claim, wherein the aerosol- generating device is a dielectric or microwave heating aerosol-generating device. Examples will now be further described with reference to the figures in which: - 22 - FIG.1 illustrates a schematic side view of an aerosol-generating device; FIG.2 illustrates a schematic upper end view of the aerosol-generating device of FIG.1; FIGS.3A-3D illustrates schematic views of aerosol-generating devices aerosol-forming articlea for use with the device; FIG.4 shows a method for generating a software that can be uploaded as a firmware to characterize puff volume; FIG. 5A illustrates a method for producing a machine learning model for an aerosol- generating device for generating an aerosol from an aerosol-forming article; FIG.5B shows a method for characterizing a puff volume using a machine learning model on an aerosol-generating device; FIG.6 illustrates a method of characterizing a usage session using a machine learning model; FIG.7 shows a system and computing apparatus that may be used to implement methods according to an example embodiment; FIG.8 illustrates a data collection scheme for training data; FIG.9 shows a system for creating training data using human puffs; and FIG.10 illustrates a change in power during a puff. FIGS.1 to 3D show different views of an exemplary aerosol-generating device 10 that can be a hand-held and portable aerosol-generating device, and has an elongate shape defined by a housing 20 that can have a substantially circularly cylindrical in form, for example as a rod or stick-like shape. The aerosol-generating device 10 comprises a heating cavity 25 located at a proximal end 21 of the housing 20 for removably receiving an aerosol-forming article 30 via an opening, the aerosol-forming article 30 comprising an aerosol-forming substrate 31. The aerosol- generating device 10 and one or more of the aerosol-forming articles 30 together form an aerosol- generating system. The aerosol-generating device 10 further comprises a battery or other power source (not shown) located within the housing 20 of the device, and an electrically operated heater 40 arranged to heat at least an aerosol-forming substrate portion 31 of an aerosol-forming article 30 when the aerosol-forming article 30 is received in the heating cavity 25. - 23 - The aerosol-generating device is configured to removably receive the aerosol-forming article 30. In the variant shown, the aerosol-forming article 30 is in the form of a cylindrical rod and comprises an aerosol-forming substrate 31. In the variant shown, the aerosol-forming substrate 31 is a solid aerosol-forming substrate comprising tobacco or a tobacco substitute. The aerosol-forming article 30 can further include a mouthpiece such as a filter 32 arranged in coaxial alignment with the aerosol-forming substrate 31 within the cylindrical rod or stick. The aerosol- forming article 30 has a diameter substantially equal to the diameter of the heating cavity 25 of the device 10 and can have a length longer than a depth of the heating cavity 25, such that when the article 30 is received in the cavity 25 of the device 10, the mouthpiece 32 extends out of the cavity 25 and may be drawn on by a user, similarly to a conventional cigarette. In use, a user inserts the article 30 into the heating cavity 25 of the aerosol-generating device 10 and turns on the device 10 by pressing a user button 50 to activate the heater 40 to start a usage session. After activation, the temperature of the heater 40 increases from an ambient temperature to a predetermined temperature for heating the aerosol-forming substrate 31 of article 30. The heater 40 heats the aerosol-forming article of the article 30 such that volatile compounds of the aerosol-forming substrate 31 can be released to form an aerosol. During a puff, the user draws on the mouthpiece of the article 30 and inhales the aerosol generated from the heated aerosol-forming substrate 31. Control electronics of the device 10 supply power to the heater 40 from the battery or other power source to maintain the temperature of the heater at an approximately constant level or to follow a specific heating profile, as a user takes several puffs on the aerosol-forming article 30. The heater continues to heat the aerosol-forming article until an end of the usage session, when the heater is deactivated and cools. In some specific examples the heater 40 may be a resistance heater. In some specific examples the heater 40 can include a coil that is fed with an alternative voltage of current, used to heat a susceptor arranged in contact with the substrate 31 within a fluctuating magnetic field generated by the coil such that it is heated by induction. Other types of heaters may also be used, or a combination of different types of heaters. FIGS.3B and 3C show example air flow channels 60, 70 that are created when the article 30 has been inserted into the heating cavity 25. The temperature may be monitored along the airflow channels 60, 70 in a location before the article 30. In an example, the temperature may be measured in a location along the airflow channel 60, 70 proximate and directly before the article 30 as shown in example temperature locations 80 and 90, for example on a side wall of the aerosol-generating device 10. While measuring the temperature directly before the article 30 may produce the most accurate results without significantly changing the device, it is to be understood - 24 - that the temperature may be measured at any location along the airflow channel. In some examples, the temperature is measured at multiple locations along the airflow channel or in locations other than the airflow channel. The location of the thermistor may be limited by the specific device configuration. FIG. 3C includes an upstream air inlet 75 where air enters the aerosol-generating device 10 and travels along the airflow channel 70 to the article 30. In the example shown in FIG.3B, the heater 40 is formed as a blade that can penetrate into the aerosol-forming substrate 31 of the article 30 for heating the substrate 31, but other types of heaters can be used. For example, FIG.3D shows a device using an inductive heater with a heating coils 34 and a susceptor 33 that is within or outside of the substrate 31. The heater 40 includes the heating coils 34 and the susceptor 33. It is to be understood that external resistive heaters, air convection heaters, dielectric / microwave heaters, or a combination of any one of these heaters may be used. At the end of the usage session, the article 30 is removed from the device 10 for disposal. Device 10 may be coupled to an external power source for charging of the battery or other power source of the device 10. The aerosol-forming article 30 for use with the device has a finite quantity of aerosol-forming substrate 31 and, thus, a usage session needs to have a finite duration to prevent a user trying to produce aerosol when the aerosol-forming substrate 31 has been depleted from at least aerosol-forming material. A usage session can be configured to have a fixed maximum duration determined by a period of time from the start of the usage session until the end of the usage session. A usage session can also configure to have a duration of less than the maximum duration if a user interaction parameter recorded during the usage session reaches a threshold before the maximum duration, for example by measuring the evolving time by a timer. In a specific embodiment, the user interaction parameter is representative of cumulative volume of aerosol generated by the one or more puffs taken by the user during the usage session. As a non-limiting and explanatory example the aerosol-generating device can be configured such that each usage session has an exemplary and non-limiting maximum duration of 6 minutes from initiation of the usage session, or a total of an exemplary and non-limiting volume of 660 ml of aerosol generated by the user (for example, equivalent to 12 puffs of 55 ml) if 660 ml of aerosol is generated within 6 minutes from initiation of the usage session. Thus, a user making a high number of short puffs, or gentle puffs, may receive a similar maximum amount of aerosol as a user taking fewer long puffs or energetic puffs. According to other aspects of the invention, an aerosol-generating device, method, and system is presented that allows to tailor the length of the user session based on the puffing style of the user, without the need of complicated sensing technology. This is accomplished by - 25 - allocating a budget aerosol volume for each user session and calculating for each puff the volume of the aerosol generated: once the total volume allowed has been reached, the user session comes to an end. The computation of the aerosol volume is performed based on a machine learning model. FIG.4 shows a method for generating a data structure as a software code or firmware to characterize puff volume or firmware generation for puff volume prediction. In a first phase, a plurality of data sets is stored for a number of different volume buckets (the higher the number of volume buckets, the higher will be the accuracy of the predictions). Data is collected for each volume bucket 410. Each data set is collected for different operating conditions and scenarios. For example, a machine learning model may be trained using consumables with different AI, consumables with different cast leaf agglomeration, consumables with different humidity level, or devices with different battery status. Data for each puff includes all possible parameters which can be measured during the puff, for example, the parameters may include the voltage, the current, the power, temperature of the thermistor (used as puff sensor). To create training data, a precise volume is known for all of the different environmental conditions. The machine learning model is trained based on the fact that data varies (because of the difference scenarios used) but they all return puffs characterized by the same aerosol volume 420. The aerosol volume may be known precisely because the puff may be carried out by a smoking machine in a controlled manner. Using the training data, the machine learning algorithm is capable of establishing correlations between the data which are used during real device operations to make puff volume predictions. Once the training data is collected, it is then fed to the machine learning algorithm to be trained and to ultimately return a “trained” firmware 430 capable of, during operation of the device, receive as input values samples during the puff (such as current, voltage and the like) and return in output a prediction of the aerosol volume of the puff. The “trained” firmware 430 is loaded into the aerosol-generating device. FIG.5A illustrates a method 500 for producing a machine learning model for an aerosol- generating device for generating an aerosol from an aerosol-forming article. The method 500 includes training 510 the machine learning model using a plurality of known target variable values over a range of operating conditions associated with one or more parameters of the aerosol- generating device. The trained machine learning model is then transferred to the aerosol- generating device 520. For example, the machine learning model is trained 510 using data collected from aerosol-generating devices with a plurality of different properties and a plurality of different operating conditions. The plurality of known target variable values are known puff - 26 - volumes and the current target variable can be a current puff volume of the aerosol-generating device. The plurality of different properties includes one or more of different aerosol-forming articles, different operating power, different humidity levels, different environmental pressure levels, different temperature levels, or different aerosol-forming article states, different types of aerosol-forming articles, for example. Once the machine learning model is trained by step 510, data of the machine learning model can be uploaded to the aerosol-generating device 10 for operation. This can be done via a communication interface, for example but not-limited via a USB-C data interface, where data of the machine learning module is uploaded as a firmware, or as a data structure that can be accessed by the existing firmware, to a memory device of the aerosol-generating device 10. The machine learning model can thereafter be used by the controller of aerosol-generating device for operation. Uploaded to the memory device of the aerosol-generating device, the machine learning model can be configured to receive as input one or more measured parameter values of the aerosol-generating device associated with current operating conditions and output a current target variable during the usage session. The current target variable may be a current puff volume, for example. The machine learning model can be based on different machine learning networks or structures, and can comprises one or more of a neural network, for example but not limited to an Artificial Neural Network (ANN), Convolutional Neural Network (CNN), Convolutional Recurrent Neural Network (CRNN) and Recurrent Neural Network (RNN), k-nearest neighbors algorithm (k-NN), support vector machines (SVM), a linear regression model, a decision tree, a decision forest, a random forest (RF), multi-layer perceptrons (MLP) fuzzy logic, correlation model, a classification model, and a clustering model. During training 510 of the machine learning model, it is possible that different types and architectures of machine learning models are trained, and the resulting trained machine learning models are benchmarked, classified, or ranked with a test data set, so that a ranking for the different trained machine learning models is generated for comparison purposes. In this respect, the method of training 510 can include a sub-step where results of the different machine learning models are displayed as a ranking, for example based on different ranking criteria. This sub-step can be such that a list is displayed in a window or area of a graphical user interface (GUI) of a display device of a computer. This allows the operator or user of the training method to select one of the different machine learning models for implementation with an aerosol-generating device 10. In this respect, the different machine learning models can be tested by inference of the trained machine learning model with benchmark or test data, to be ranked with a ranking value, for example a percentage value of errors or deviation from an actual value of the target variable, for - 27 - example, a cumulative puff volume during a usage session. Thereby, it is possible to choose the best suited trained machine learning model for upload and operation with the aerosol-generating device 10 for operationalization. For example, a purely result-oriented comparison can be done related to the estimated or calculated one or more target variables, where the performance of the test or benchmark data for the specific target variable is used. For example, a trained machine learning model could be chosen based on the best accuracy of the estimation or calculation of the puff-related characteristics, such as puff volume. This could be a statistical variation of a puff- by-puff basis, or for a cumulative puff volume per usage session. However, it is also possible to take other results into account, for example use of memory space used by the uploaded and trained machine learning model, power consumption of the trained machine learning model, latency time duration of the estimation or calculation by the trained machine learning model, potential false result generation of the trained machine learning model, for example caused by local maxima / minima. FIG.5B shows a method 550 for characterizing a puff volume using a machine learning model on an aerosol-generating device 10. One or more parameters associated with current operating conditions of the aerosol-generating device are monitored 560 during operation of the aerosol-generating device. The one or more parameters associated with the aerosol device include one or more of a voltage, a current, a power, and a temperature associated with the aerosol-generating device. For example, the temperature of the puff sensor, and a temperature of the PCB of the aerosol-generating device. A machine learning model is used 570 to determine a current target variable based on the one or more parameters. The current target variable may control operation of the aerosol- generating device. For example, the current target variable may be used to end a usage session of the aerosol-generating device, for example, stopping or ramping down the heating. While examples herein describe the current target variable as being a puff volume or a puff detection, it is to be understood that the current target variables may include one or more other target variables associated with the aerosol-generating device, for example other parameters related to the usage session and the one or more puffs taken by the user. FIG.6 illustrates a method of characterizing a usage session. A user inserts an aerosol- forming article into the aerosol-generating device and initiates a usage session by actuating the user button 50. This indicates the start of the usage session 601. Power is supplied from a battery in the aerosol-generating device to the heater 40 until the temperature of the aerosol-forming substrate 31 reaches a predetermined operating temperature. This temperature may be, for example, about 330 degrees Centigrade. - 28 - Parameters associated with the aerosol-generating device are monitored 602. In an example, the power signal is monitored 602. When the user puffs 603, the heater is cooled because of the cooling effect of the airflow, as environmental and relative cold air is drawn into the heated substrate 31 by the user. Thus, the power that needs to be supplied to the heater to maintain the operating temperature increases. The power supplied increases and the correct temperature is maintained. The power signal may be determined by using the machine learning model and the one or more sensors described herein. The presence of a user puff is detected 604. For example, the user puff may be detected by analyzing the power signal or using the machine learning model. A puff start point and a puff end point are determined by means of this analysis. The volume of aerosol generated during the puff is calculated 606 using the machine learning model. The puff volume is added to a cumulative total of volume generated during the usage session. It is determined 607 whether a cumulative puff volume equals or exceeds a predefined maximum allowed (threshold) volume. If the cumulative total volume equals or exceeds the predetermined maximum permissible aerosol volume for the usage session (for example 660 ml), the usage session is ended 608. If the cumulative total volume does not equal or exceeds the predetermined maximum permissible aerosol volume for the usage session then the session remains active and the user may take another puff 602. The usage session remains active until the user has generated the maximum permissible aerosol volume or until a maximum time threshold is reached, for example. The methods and processes described above can be implemented on computer hardware. In FIG.7, a block diagram shows a system and computing apparatus 700 that may be used to implement methods according to an example embodiment (for example, as a computer, a mobile device, a smart sensor, a control system, and the like.). The components may be implemented as integrated circuits (Ics), portions thereof, discrete electronic devices, or other modules, instruction sets, programmable logic or algorithms, hardware, hardware accelerators, software, firmware, or a combination thereof. One or more sensors 712, 713, 714 are disposed on the aerosol-generating apparatus. The sensors may include one or more of a power sensor, a voltage sensor, a current sensor, one or more temperature sensors (for example, a cavity temperature sensor and a PCB temperature sensor), and an accelerometer or a position sensor. In this respect the aerosol-generating device 10 can be equipped with an accelerometer or a position sensor, for example an inertial measurement unit (IMU), that can measure movements, orientation, and accelerations to the - 29 - aerosol-generating device 10 while the user is holding the device during a usage session, or while taking one or more puffs. With data from the accelerometer, it is possible that false or ghost puffs are detected and eliminated from the puff detection and puff volume. In this respect, some types of motions that are departed to the aerosol-generating device 10 could cause a ghost puff or an air motion that is not intended or done by the user, and data patterns from the accelerometer could provide for valid data to exclude these puffs. Therefore, data of such sensor can be provided as input data to the trained machine learning network for the estimation or calculation of the one or more target variables. It is also that certain types of unreliable or expensive sensors are avoided, or its data is not used. For example, the one or more sensors or plurality of sensors may not include any sensors such as humidity sensors that capture a humidity level from the external environment, as these types of sensors may be easily covered or otherwise obstructed for reliable measurements, or other sensors that can be clogged or otherwise be rendered inoperative due to particle accumulation and other environmental issues, for example flow sensors. The sensor data may be fed into controller 720. The controller 720 may include a user interface 725 conventional computing hardware such as a central processor 721, memory 722, input / output (I / O) interfaces 723, and a non-volatile data storage unit 724 (for example, hard disk drives, solid state drives). The processor 721 may include any one or more of a microprocessor, a controller, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or equivalent discrete or integrated logic circuitry. In some embodiments, the processor 721 may include multiple components, such as any combination of one or more microprocessors, one or more controllers, one or more DSPs, one or more ASICs, or one or more FPGAs, as well as other discrete or integrated logic circuitry. The functions attributed to the controller 720 or processor 721 herein may be embodied as software, firmware, hardware, or any combination of these. Certain functionality of the controller 720 may also be performed in the cloud or other distributed computing systems operably connected to the processor 721. The memory 722 may include any volatile, non-volatile, magnetic, optical, or electrical media, such as a random-access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), electrically erasable programmable ROM (EEPROM), flash memory, or any other digital media. While shown as both being incorporated into the controller 720, the memory 722 and the processor 721 could be contained in separate modules. The controller 720 includes an external data interface 726 that receives data from the sensors 712, 713, 714. The data storage unit 724 stores a machine learning model 728 configured to predict current target variables based on device parameters sensed via sensors 712, 713, 714. - 30 - Usage session management 732 may be used to determine a usage session and puff characterization 730 based on an output from the machine learning model. For example, usage session management may determine that a usage session should be ended based on a cumulative puff volume reaching a predefined threshold. FIG.8 illustrates a data collection scheme for training data. In this example, three different volume buckets are shown, but it is to be understood that more or fewer known volume buckets may be used for the training data. Each of the volume buckets determines various parameters for four different scenarios. It is to be understood, the more or less than four scenarios may be used. Table 1 shows predictions versus measurements for puff volume using the machine learning model. Table 1 Puff Volume (ml) Number of 35ml Prediction 55ml Prediction 70ml Prediction Candidates 35 20339 84% 8% 8% 55 26541 6% 86% 8% 70 27197 6% 10% 84% In this example, the accuracy for the 35ML and the 70ML volume buckets was 84% and the accuracy rating of the 55ML puff volume was 86%. In this example, more than 20k candidates were used to train the model. The candidates are split 80 % for training and 20 % for testing. The prediction result is the result of correct identification of classes on the 20% test candidates. The data was generated by using a classifier model trained for three (3) classes. First, the data was split into two (2) buckets, with one bucket for training, and the other one for testing. This can be done automatically by a machine learning tool, for example a commercially available machine learning tool that can operate on a commercially available computer, such as a PC or a Macintosh computer. The machine learning tool then generates a first model. This first model is then run on the computer and can be inputted with test data set or a benchmark data set for so that the performance or accuracy of the first model can be evaluated. The first model running on the computer can output the score / accuracy of test data or benchmark data set. The accuracy numbers, for example, in percentage of error, are represented in Table 1. Next, one or more additional models can be tested with the test or benchmark data set, to see the score / accuracy data for comparison and selection purposes. - 31 - FIG.9 illustrates a data collection method for generating a data structure as a software code or firmware to characterize puff volume using human puffs by a single user or a plurality of users of an aerosol-generating device 910. The method can be performed instead of using the smoking or inhalation machine in a controlled manner to generate the airflow, or can be performed by using both the smoking or inhalation machine. The human puffs are taken during a high number of usage sessions with the aerosol-generating device 910. The puffs taken by human inhalation are measured and characterized accurately by using a reference flow meter 920 that provides flow measurement data 930. The reference flow meter is fluidically arranged between a mouth of the user and the aerosol-generating device 910. In an example, the reference flow meter 920 is disposed on a mouthpiece side of the aerosol-generating device 910. This flow measurement data 930 may be used to provide a reference value 940 for the puff volume for each puff, instead of using the airflow generated by the smoking or inhalation machine, or reference values can be generated by both human inhalation and smoking or inhalation machine. For each puff various other internal sensor data 950 is obtained from the aerosol-generating device 910. For example, one or more of a power sensor, a voltage sensor, a current sensor, one or more temperature sensors (for example, a cavity temperature sensor and a PCB temperature sensor), and an accelerometer. The sensor data 950 and the reference values 940 may be used to generate training data for use with one or more different machine learning algorithms to generate one or more trained machine learning models 960, 970, 980. The trained models 960, 970, 980 may include one or more of an aerosol-generating device simulator, a model generated with a classifier and a model generated with a classifier with a low memory footprint. In a variant, instead of performing a data collection method for generating a data structure to characterize puff volume, it is possible that other types of known target variables are used to train the machine learning model. For example, it is possible to train a machine learning model for detecting whether the aerosol is hot (exceeds a certain threshold temperature) or not while a puff is taken. With such trained machine learning model, it is possible to provide for information to the aerosol-generating device 10 that will allow to mitigate against the hot aerosol effect. In such system that performs the data collection method, a reference temperature sensor can be arranged between the mouthpiece of the aerosol-generating article and the smoking or inhalation machine (or human user), that can be used to generate a reference temperature value during a plurality of puffs, serving as an example of a known target variable value. This data can be used for training the machine learning model, such that the trained and operationalized machine learning model can infer a temperature of the aerosol that exits from the mouthpiece. As another example, the machine learning model can be trained by a human user, where the identity of the user serves as a known target variable, - 32 - FIG.10 shows that the energy strictly associated to a puff may be calculated as the integral calculus of the power signal during the puff, minus the energy that would be spent anyway even without a puff, as indicated in [1]. [1] Similarly, it is also possible to correlate the power to the air flow, which equals the volume per time unit. One or more of the sensors described herein may be used to determine the data indicative of a change in heating power as a result of air cooling. The usage session has a maximum permissible volume of aerosol to be delivered. Every puff contributes to the maximum permissible volume. Once the threshold has been reached, the experience ends. Therefore, the experience is not tied to a predetermined number of puffs, but to the way the user actually puffs on the device. According to another aspect of the present invention, it is possible to use data gathered from the aerosol-generating device during the one or more usage sessions by a specific user can be used to further improve the machine learning model. For example, according to an aspect of the present invention, a method for improving a machine learning model can be provided, where first a user uses his aerosol-generating device, so that data on the usage sessions can be recorded. For example, data from one or more sensors can be recorded to a memory device during one or more usage session from a user, for example power data on the power consumption of the heater, and temperature data from the one or more temperature sensors. In another variant, data from an acceleration sensor, for example an IMU can be recorded. In addition, the corresponding generated or estimated output data of the machine learning model can be recorded to a memory device, which includes a current target variables that will be based on the recorded one or more parameters, to thereby create a user- specific data set of the usage and results of the currently uploaded and operating machine learning model. Next, in another step, a data communication can be established with an external data processing device, to download the user-specific data set that has been generated by the trained machine learning model and provided to a computing environment to further improve the trained machine learning model, by using the user-specific data set. The downloaded data set can be used to correct an error in the calculation or estimation, For the purpose of the present description and of the appended claims, except where otherwise indicated, all numbers expressing amounts, quantities, percentages, and so forth, are to be understood as being modified in all instances by the term “about.” Also, all ranges include the maximum and minimum points disclosed and include any intermediate ranges therein, which - 33 - may or may not be specifically enumerated herein. In this context, therefore, a number A is understood as A ± 2% of A. Within this context, a number A may be considered to include numerical values that are within general standard error for the measurement of the property that the number A modifies. The number A, in some instances as used in the appended claims, may deviate by the percentages enumerated above provided that the amount by which A deviates does not materially affect the basic and novel characteristic(s) of the claimed invention. Also, all ranges include the maximum and minimum points disclosed and include any intermediate ranges therein, which may or may not be specifically enumerated herein.

Claims

- 34 - CLAIMS 1. An aerosol-generating device for generating an aerosol from an aerosol-forming article having a solid aerosol-forming substrate, the aerosol-generating device comprising: a heater for heating the solid aerosol-forming substrate of the aerosol-forming article during a usage session, the usage session having a duration allowing for taking several puffs; a power supply for supplying power to the heater to generate the aerosol; one or more sensors for monitoring one or more parameters associated with current operating conditions of the aerosol-generating device during the usage session of the aerosol- generating device; a storage device for storing a machine learning model trained using a plurality of known target variable values over a range of operating conditions associated with the one or more parameters of the aerosol-generating device; and a controller coupled to the power supply, the one or more sensors, and the storage device, the controller for using the machine learning model to determine a current target variable related to a puff based on the one or more parameters.

2. The aerosol-generating device of claim 1, wherein the one or more sensors comprises one or more of a power sensor, one or more temperature sensors, and an accelerometer.

3. The aerosol-generating device of claim 1, wherein the aerosol-forming article comprises an aerosol-forming substrate and the aerosol-generating device comprises heater configured to heat the aerosol-forming substrate, wherein the one or more sensors comprises one or both of a sensor that can provide for data that is indicative of a temperature caused by a cooling effect of an airflow to an element that forms an airflow path and a sensor that can provide for data that is indicative of a power consumption of the heater.

4. The aerosol-generating device of claim 3, wherein the aerosol-forming article comprises an aerosol-forming substrate, wherein the sensor that provides data indicative of the temperature includes a temperature sensor that can measure a temperature of a side wall of the airflow path upstream of the aerosol-forming substrate.- 35 - 5. The aerosol-generating device of claim 3, wherein the sensor that provides data indicative of the power consumption of the heater includes one or both of a voltage and a current sensor.

6. The aerosol-generating device of claim 1, wherein the current target variable controls operation of the aerosol-generating device, wherein the current target variable includes data that characterizes one or more puffs taken by a user of the aerosol-generating device.

7. The aerosol-generating device of claim 1, wherein the aerosol-generating device generates aerosol during a usage session, wherein the controller: determines a start of the usage session; monitors the one or more parameters associated with current operating conditions of the aerosol-generating device during the usage session, and uses the current target variable to determine an end of the usage session.

8. The aerosol-generating device of claim 7, wherein the plurality of known target variable values are known puff volumes and the current target variable is a current puff volume of the aerosol-generating device, and wherein using the current target variable to determine the end of the usage session comprises: determining a cumulative puff volume measured from the start of the usage session; and using the cumulative puff volume to determine the end of the usage session.

9. The aerosol-generating device of claim 8, wherein the controller ends the usage session when the cumulative puff volume reaches a predefined threshold.

10. The aerosol-generating device of claim 1, wherein the aerosol-forming article is a heat- not-burn aerosol-forming article.

11. A method of operating an aerosol-generating device, comprising: monitoring one or more parameters associated with current operating conditions of the aerosol-generating device during operation of the aerosol-generating device; and- 36 - using a machine learning model to determine a current target variable based on the one or more parameters, the machine learning model trained using a plurality of known target variable values over a range of operating conditions associated with the one or more parameters of the aerosol-generating device.

12. A method for producing a machine learning model for an aerosol-generating device having a solid aerosol-forming substrate for generating an aerosol from an aerosol-forming article, the method comprising: training the machine learning model using a plurality of known target variable values over a range of operating conditions associated with one or more parameters of the aerosol- generating device; and transferring the machine learning model to a controller of the aerosol-generating device, the machine learning model configured to receive as input one or more measured parameter values of the aerosol-generating device associated with current operating conditions and output a current target variable.

13. The method of claim 12, wherein the range of operating conditions comprise different aerosol-forming articles, different operating power, different environmental humidity levels, different environmental pressure levels, different environmental temperature levels, and different aerosol-forming article states.

14. The method of claim 12, wherein the plurality of known target variable values are known puff volumes and the current target variable is at least one of current puff volume of the aerosol- generating device, a temperature of aerosol exiting the aerosol-forming article for consumption.

15. The method of claim 12, wherein the one or more parameters associated with the aerosol-generating device comprise one or more of a voltage, a current, a power, and a temperature of an airflow channel of the aerosol-generating device.

Images